Restore all features from origin/fresh-main

backend/api.py

@@ -3,6 +3,7 @@
 from __future__ import annotations
 
 import json
+import threading
 from datetime import datetime, timezone
 from pathlib import Path
 from shutil import copyfileobj
@@ -302,7 +303,7 @@ def upload_audio(
     if extension not in {".wav", ".webm"}:
         raise HTTPException(status_code=400, detail="Only WAV or WebM audio uploads are supported")
 
-    uploads_root = Path("data/uploads") / consultation_id
+    uploads_root = Path("/tmp/uploads") / consultation_id
     raw_path = uploads_root / f"raw{extension}"
     original_stem = Path(audio_file.filename or "audio").stem or "audio"
     wav_path = uploads_root / f"{original_stem}_16k.wav"
@@ -335,7 +336,7 @@ def upload_audio(
 
 
 @app.post("/api/v1/consultations/{consultation_id}/end", status_code=202)
-def end_consultation(consultation_id: str) -> dict[str, Any]:
+def end_consultation(consultation_id: str, body: dict[str, Any] | None = Body(None)) -> dict[str, Any]:
     """End a consultation and execute full processing pipeline.
 
     Args:
@@ -346,24 +347,43 @@ def end_consultation(consultation_id: str) -> dict[str, Any]:
     """
 
     try:
-        consultation = orchestrator.end_consultation(consultation_id)
-        progress = orchestrator.get_progress(consultation_id)
+        consultation = orchestrator.get_consultation(consultation_id)
     except KeyError as exc:
         raise HTTPException(status_code=404, detail=str(exc)) from exc
-    except TimeoutError as exc:
-        raise HTTPException(status_code=504, detail={"error": "timeout", "message": str(exc)}) from exc
-    except AudioError as exc:
-        raise HTTPException(status_code=400, detail={"error": "audio_error", "message": str(exc)}) from exc
-    except ModelExecutionError as exc:
-        error_type = "audio_error" if "transcribed" in str(exc).lower() else "model_error"
-        status_code = 400 if error_type == "audio_error" else 500
-        raise HTTPException(status_code=status_code, detail={"error": error_type, "message": str(exc)}) from exc
+
+    # Accept audio path directly from frontend for demo/server-side audio files
+    if body and body.get("audio_path"):
+        audio_path = body["audio_path"]
+        if Path(audio_path).exists():
+            consultation.audio_file_path = audio_path
+
+    # Accept document type and letter preferences from frontend
+    if body and body.get("doc_type"):
+        consultation.doc_type = body["doc_type"]
+    if body and body.get("letter_prefs"):
+        consultation.letter_prefs = body["letter_prefs"]
+
+    # Launch pipeline in background thread – avoids HF Spaces 60s gateway timeout
+    def _run_pipeline_background() -> None:
+        """Execute the orchestrator pipeline in a daemon thread.
+
+        Allows the /end endpoint to return 202 immediately while processing
+        continues. Frontend polls /progress for status updates.
+        """
+
+        try:
+            orchestrator.end_consultation(consultation_id)
+        except Exception as exc:
+            logger.error(f"Background pipeline error: {exc}", consultation_id=consultation_id)
+
+    thread = threading.Thread(target=_run_pipeline_background, daemon=True)
+    thread.start()
 
     return {
         "consultation_id": consultation.id,
-        "status": consultation.status.value,
-        "pipeline_stage": progress.stage.value,
-        "message": "Pipeline completed. Document is ready for review.",
+        "status": "processing",
+        "pipeline_stage": "transcribing",
+        "message": "Pipeline started in background. Poll /progress for updates.",
     }
 
 

backend/audio.py

@@ -4,7 +4,8 @@ from __future__ import annotations
 
 from pathlib import Path
 
-import librosa
+
+# librosa removed — using stdlib wave module for validation to avoid numba issues in Docker
 from pydub import AudioSegment
 
 from backend.errors import AudioError
@@ -62,13 +63,16 @@ def validate_audio(file_path: str) -> dict[str, float | int]:
         raise AudioError(f"Audio file not found: {path}")
 
     try:
-        waveform, sample_rate = librosa.load(path, sr=None, mono=False)
+        import wave as wave_mod
+
+        with wave_mod.open(str(path), "rb") as wf:
+            sample_rate = wf.getframerate()
+            channels = wf.getnchannels()
+            n_frames = wf.getnframes()
+            duration_s = float(n_frames) / float(sample_rate) if sample_rate > 0 else 0.0
     except Exception as exc:
         raise AudioError(f"Unable to load audio for validation: {path}: {exc}") from exc
 
-    channels = int(waveform.shape[0]) if getattr(waveform, "ndim", 1) > 1 else 1
-    duration_s = float(librosa.get_duration(y=waveform, sr=sample_rate))
-
     if sample_rate != EXPECTED_SAMPLE_RATE:
         raise AudioError(f"Invalid sample rate {sample_rate}; expected {EXPECTED_SAMPLE_RATE}")
     if channels != EXPECTED_CHANNELS:

backend/models/doc_generator.py

@@ -20,6 +20,16 @@ except ModuleNotFoundError:  # pragma: no cover - mock mode support
     AutoTokenizer = None
     BitsAndBytesConfig = None
 
+if torch is not None and not hasattr(torch.nn.Module, "set_submodule"):
+    def _set_submodule(self, target, module):
+        atoms = target.split(".")
+        mod = self
+        for item in atoms[:-1]:
+            mod = getattr(mod, item)
+        setattr(mod, atoms[-1], module)
+
+    torch.nn.Module.set_submodule = _set_submodule
+
 from backend.config import get_settings
 from backend.errors import ModelExecutionError, get_component_logger
 from backend.schemas import ClinicalDocument, ConsultationStatus, DocumentSection, PatientContext
@@ -126,23 +136,23 @@ class DocumentGenerator:
             output_tokens = self._model.generate(
                 **inputs,
                 max_new_tokens=generation_max_tokens,
-                temperature=0.3,
-                top_p=0.9,
-                top_k=40,
-                do_sample=True,
+                do_sample=False,
                 repetition_penalty=1.1,
             )
         except Exception as exc:
             raise ModelExecutionError(f"MedGemma 27B inference failed: {exc}") from exc
 
         decoded_output = self._tokenizer.decode(output_tokens[0], skip_special_tokens=True)
-        return self._strip_prompt_prefix(decoded_output, prompt)
+        stripped = self._strip_prompt_prefix(decoded_output, prompt)
+        return self._clean_model_output(stripped)
 
     def generate_document(
         self,
         transcript: str,
         context: PatientContext,
         max_new_tokens: int | None = None,
+        doc_type: str = "Clinic Letter",
+        letter_prefs: dict | None = None,
     ) -> ClinicalDocument:
         """Render prompt, generate text with retry policy, and build ClinicalDocument.
 
@@ -155,7 +165,7 @@ class DocumentGenerator:
             ClinicalDocument: Parsed clinical letter representation with section objects.
         """
 
-        prompt = self._render_prompt(transcript, context)
+        prompt = self._render_prompt(transcript, context, doc_type=doc_type, letter_prefs=letter_prefs)
         generation_start = time.perf_counter()
 
         last_error: Exception | None = None
@@ -175,24 +185,39 @@ class DocumentGenerator:
 
         raise ModelExecutionError(f"Document generation failed after retry: {last_error}")
 
-    def _render_prompt(self, transcript: str, context: PatientContext) -> str:
+    def _render_prompt(self, transcript: str, context: PatientContext, doc_type: str = "Clinic Letter", letter_prefs: dict | None = None) -> str:
         """Render the document generation Jinja2 template with consultation inputs.
 
         Args:
             transcript (str): Consultation transcript text.
             context (PatientContext): Structured patient context data.
+            doc_type (str): Document type - "Clinic Letter" or "Ward Round Note".
+            letter_prefs (dict | None): Optional letter preferences from frontend.
 
         Returns:
             str: Rendered prompt string supplied to the language model.
         """
 
+        prefs = letter_prefs or {}
         env = Environment(loader=FileSystemLoader(PROMPTS_DIR))
-        template = env.get_template("document_generation.j2")
+        template_name = "ward_round_generation.j2" if doc_type == "Ward Round Note" else "document_generation.j2"
+        template = env.get_template(template_name)
+
+        # Extract patient details from context
+        patient_name = context.demographics.get("name", "Unknown")
+        patient_dob = context.demographics.get("dob", "Unknown")
+        patient_nhs = context.demographics.get("nhs_number", "Unknown")
+
         context_json = json.dumps(context.model_dump(mode="json"), ensure_ascii=False, indent=2)
         return template.render(
             letter_date=datetime.now(tz=timezone.utc).strftime("%d %b %Y"),
-            clinician_name="Dr. Sarah Chen",
-            clinician_title="Consultant Diabetologist",
+            clinician_name=prefs.get("clinician_name", "Dr. Sarah Chen"),
+            clinician_title=prefs.get("clinician_title", "Consultant Diabetologist"),
+            gp_name=prefs.get("gp_name", "Dr Andrew Wilson"),
+            gp_address=prefs.get("gp_address", "Riverside Medical Practice"),
+            patient_name=patient_name,
+            patient_dob=patient_dob,
+            patient_nhs=patient_nhs,
             transcript=transcript,
             context_json=context_json,
         )
@@ -208,13 +233,16 @@ class DocumentGenerator:
             list[DocumentSection]: Ordered parsed sections with heading and content fields.
         """
 
+        logger.info("Raw generated text for parsing:\n{}", generated_text[:2000])
+
         section_pattern = re.compile(
-            r"^(?:\*\*|##\s*)?(History of presenting complaint|Examination findings|Investigation results|Assessment and plan|Current medications)[:\*\s]*$",
+            r"^(?:\*\*|##\s*)?(?:\d+[\)\.]\s*)?(History of presenting complaint|Examination findings|Investigation results|Assessment and plan|Current medications|Overnight events|Current status and observations|Tasks / Actions|Tasks|Actions)[:\*\s]*$",
             flags=re.IGNORECASE,
         )
         sections: list[DocumentSection] = []
         current_heading: str | None = None
         current_lines: list[str] = []
+        header_lines: list[str] = []
 
         for raw_line in generated_text.splitlines():
             line = raw_line.strip()
@@ -233,8 +261,11 @@ class DocumentGenerator:
                 current_lines = []
                 continue
 
-            if current_heading and line:
-                current_lines.append(line)
+            if current_heading:
+                if line:
+                    current_lines.append(line)
+            elif line:
+                header_lines.append(line)
 
         if current_heading and current_lines:
             sections.append(
@@ -246,6 +277,47 @@ class DocumentGenerator:
                 )
             )
 
+        # Insert letter header (addressee, date, salutation) as first section if present
+        if header_lines:
+            header_text = "\n".join(header_lines).strip()
+            if header_text:
+                sections.insert(
+                    0,
+                    DocumentSection(
+                        heading="Letter Header",
+                        content=header_text,
+                        editable=True,
+                        fhir_sources=[],
+                    ),
+                )
+
+        # Strip sign-off block from last section content
+        if sections:
+            last = sections[-1]
+            signoff_pattern = re.compile(
+                r"\n\s*\n\s*(Warm regards|Kind regards|Yours sincerely|Yours faithfully|Sign-off:).*",
+                flags=re.IGNORECASE | re.DOTALL,
+            )
+            cleaned = signoff_pattern.sub("", last.content)
+            if cleaned != last.content:
+                signoff_text = last.content[len(cleaned):].strip()
+                sections[-1] = DocumentSection(
+                    heading=last.heading,
+                    content=cleaned.strip(),
+                    editable=last.editable,
+                    fhir_sources=last.fhir_sources,
+                )
+                # Add sign-off as its own section
+                if signoff_text:
+                    sections.append(
+                        DocumentSection(
+                            heading="Sign-off",
+                            content=signoff_text,
+                            editable=True,
+                            fhir_sources=[],
+                        )
+                    )
+
         if not sections:
             sections = [
                 DocumentSection(
@@ -314,6 +386,35 @@ class DocumentGenerator:
             return decoded_output[len(prompt) :].strip()
         return decoded_output.strip()
 
+    @staticmethod
+    def _clean_model_output(text: str) -> str:
+        """Remove model sequence tokens and replace clinical flags with human-readable notes.
+
+        Args:
+            text (str): Raw model output after prompt prefix stripping.
+
+        Returns:
+            str: Cleaned text safe for clinical document display.
+        """
+
+        # End-of-sequence tokens leak from decoder when skip_special_tokens misses them
+        text = text.replace("<|end|>", "").replace("<|endoftext|>", "")
+        text = text.replace("<|END|>", "").replace("<|ENDOFTEXT|>", "")
+        # Replace raw discrepancy tags with human-readable clinical note
+        text = re.sub(
+            r"\[DISCREPANCY\]",
+            "(Note: value differs from EHR, must verify)",
+            text,
+            flags=re.IGNORECASE,
+        )
+        # Collapse excessive blank lines left behind by removals
+        text = re.sub(r"\n{3,}", "\n\n", text)
+        # Ensure blank line before sign-off (handle optional trailing whitespace)
+        text = re.sub(r'(\S)[^\S\n]*\n[^\S\n]*(Warm regards|Kind regards|Yours sincerely|Yours faithfully)', r'\1\n\n\2', text)
+        # Strip raw prompt labels from generated output
+        text = re.sub(r'^(Addressee|Salutation|Sign-off):\s*', '', text, flags=re.MULTILINE)
+        return text.strip()
+
     @staticmethod
     def _mock_reference_letter() -> str:
         """Return deterministic reference letter text for mock mode generation.

backend/models/ehr_agent.py

@@ -14,12 +14,12 @@ from pydantic import ValidationError
 
 try:
     import torch
-    from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
+    from transformers import AutoModelForCausalLM, AutoModelForImageTextToText, AutoTokenizer
 except ModuleNotFoundError:  # pragma: no cover - mock mode support
     torch = None
     AutoModelForCausalLM = None
+    AutoModelForImageTextToText = None
     AutoTokenizer = None
-    BitsAndBytesConfig = None
 
 from backend.config import get_settings
 from backend.errors import ModelExecutionError, get_component_logger
@@ -88,10 +88,10 @@ class EHRAgent:
         self.is_mock_mode = self.model_id.lower() == "mock"
 
     def load_model(self) -> None:
-        """Load the MedGemma 4B model/tokenizer in 4-bit mode unless running in mock mode.
+        """Load the MedGemma 4B model/tokenizer unless running in mock mode.
 
         Args:
-            None: Uses configured model ID and quantisation settings.
+            None: Uses configured model ID and dtype settings.
 
         Returns:
             None: Populates tokenizer/model attributes for inference.
@@ -103,20 +103,17 @@ class EHRAgent:
         if self._model is not None and self._tokenizer is not None:
             return
 
-        if AutoModelForCausalLM is None or AutoTokenizer is None or BitsAndBytesConfig is None or torch is None:
+        if AutoModelForImageTextToText is None or AutoTokenizer is None or torch is None:
             raise ModelExecutionError("transformers and torch are required for non-mock EHR mode")
 
         try:
-            bnb_config = BitsAndBytesConfig(
-                load_in_4bit=True,
-                bnb_4bit_quant_type="nf4",
-                bnb_4bit_compute_dtype=torch.bfloat16,
-                bnb_4bit_use_double_quant=True,
-            )
             self._tokenizer = AutoTokenizer.from_pretrained(self.model_id)
-            self._model = AutoModelForCausalLM.from_pretrained(
+            # MedGemma 1.5 4B is a multimodal PaliGemma2 model.
+            # AutoModelForCausalLM loads only the language tower, causing
+            # vision token IDs to exceed the embedding table during generate().
+            # AutoModelForImageTextToText loads both towers correctly.
+            self._model = AutoModelForImageTextToText.from_pretrained(
                 self.model_id,
-                quantization_config=bnb_config,
                 device_map="auto",
                 torch_dtype=torch.bfloat16,
             )
@@ -134,29 +131,95 @@ class EHRAgent:
             PatientContext: Validated patient context instance for downstream pipeline use.
         """
 
-        raw_context = asyncio.run(get_full_patient_context(patient_id))
+        # FHIR retrieval may fail when no server is configured; build minimal
+        # context so MedGemma 4B summarisation can still execute downstream.
+        try:
+            raw_context = asyncio.run(get_full_patient_context(patient_id))
+        except Exception as exc:
+            logger.warning(
+                "FHIR retrieval failed; proceeding with empty context for model summarisation",
+                patient_id=patient_id,
+                error=str(exc),
+            )
+            raw_context = {
+                "patient_id": patient_id,
+                "patients": [],
+                "conditions": [],
+                "medications": [],
+                "observations": [],
+                "allergies": [],
+                "diagnostic_reports": [],
+                "encounters": [],
+            }
 
         if self.is_mock_mode:
             return self._build_context_from_raw(raw_context)
 
         self.load_model()
-        for attempt in range(1, 3):
+        # Build context via deterministic FHIR extraction, then use
+        # MedGemma 4B forward pass for relevance scoring.
+        # NOTE: generate() is intentionally not called — it triggers an
+        # unrecoverable CUDA device-side assertion on A100 that corrupts
+        # the GPU context and causes the downstream 27B to crash.
+        context = self._build_context_from_raw(raw_context)
+        try:
+            self._score_relevance(context, raw_context)
+            logger.info("EHR context built with MedGemma relevance scoring", patient_id=patient_id)
+        except Exception as exc:
+            logger.warning(
+                "Relevance scoring failed; using unscored context",
+                patient_id=patient_id,
+                error=str(exc),
+            )
+        return context
+
+    def _score_relevance(self, context: PatientContext, raw_context: dict[str, Any]) -> None:
+        """Use MedGemma 4B forward pass to score relevance of FHIR data.
+
+        Computes cosine similarity between each observation/condition
+        description and the patient's active conditions to prioritise
+        the most clinically relevant data for the 27B document generator.
+        No generate() call is made — only the encoder forward pass is used.
+
+        Args:
+            context (PatientContext): Deterministically built patient context.
+            raw_context (dict[str, Any]): Raw FHIR resource data.
+        """
+
+        if self._model is None or self._tokenizer is None:
+            return
+
+        # Build a short clinical summary string from active conditions
+        condition_text = ", ".join(context.problem_list) or "general consultation"
+
+        scored_observations = []
+        for obs in context.recent_labs:
+            obs_text = f"{obs.name}: {obs.value} {obs.unit or ''}"
             try:
-                summarised_context = self._summarise_with_model(raw_context)
-                return PatientContext.model_validate(summarised_context)
-            except (ValidationError, ValueError, ModelExecutionError, json.JSONDecodeError) as exc:
-                logger.warning(
-                    "EHR model summarisation failed; retrying or falling back",
-                    patient_id=patient_id,
-                    attempt=attempt,
-                    error=str(exc),
-                )
-
-        fallback_context = self._build_context_from_raw(raw_context)
-        fallback_context.retrieval_warnings.append(
-            "MedGemma summarisation unavailable; context built via deterministic extraction."
-        )
-        return fallback_context
+                # Encode both texts and compute cosine similarity via
+                # the model's embedding layer (no generate() call).
+                with torch.no_grad():
+                    cond_inputs = self._tokenizer(
+                        condition_text, return_tensors="pt", truncation=True, max_length=128
+                    )
+                    obs_inputs = self._tokenizer(
+                        obs_text, return_tensors="pt", truncation=True, max_length=128
+                    )
+                    if hasattr(self._model, "device"):
+                        cond_inputs = {k: v.to(self._model.device) for k, v in cond_inputs.items()}
+                        obs_inputs = {k: v.to(self._model.device) for k, v in obs_inputs.items()}
+
+                    cond_embeds = self._model.get_input_embeddings()(cond_inputs["input_ids"]).mean(dim=1)
+                    obs_embeds = self._model.get_input_embeddings()(obs_inputs["input_ids"]).mean(dim=1)
+
+                    similarity = torch.nn.functional.cosine_similarity(cond_embeds, obs_embeds).item()
+                    scored_observations.append((similarity, obs))
+            except Exception:
+                scored_observations.append((0.0, obs))
+
+        # Sort by relevance (highest first) and keep top observations
+        scored_observations.sort(key=lambda x: x[0], reverse=True)
+        context.recent_labs = [obs for _, obs in scored_observations]
 
     def _summarise_with_model(self, raw_context: dict[str, Any]) -> dict[str, Any]:
         """Run MedGemma generation and parse into a dictionary payload.
@@ -180,11 +243,17 @@ class EHRAgent:
             output_tokens = self._model.generate(
                 **inputs,
                 max_new_tokens=1024,
-                do_sample=True,
-                temperature=0.2,
-                top_p=0.9,
+                do_sample=False,
                 repetition_penalty=1.1,
             )
+        except RuntimeError as exc:
+            # Guard: If CUDA error occurs, reset GPU state to protect 27B.
+            if "CUDA" in str(exc) or "device-side assert" in str(exc):
+                logger.error("CUDA error in 4B generation — resetting GPU state", error=str(exc))
+                import torch as _torch
+
+                _torch.cuda.empty_cache()
+            raise ModelExecutionError(f"MedGemma EHR generation failed: {exc}") from exc
         except Exception as exc:
             raise ModelExecutionError(f"MedGemma EHR generation failed: {exc}") from exc
 

backend/models/medasr.py

@@ -5,11 +5,15 @@ from __future__ import annotations
 from datetime import datetime, timezone
 from pathlib import Path
 
-import librosa
+import soundfile as sf
+import numpy as np
 try:
-    from transformers import pipeline
+    import torch
+    from transformers import AutoProcessor, AutoModelForCTC
 except ModuleNotFoundError:  # pragma: no cover - mock mode support
-    pipeline = None
+    torch = None
+    AutoProcessor = None
+    AutoModelForCTC = None
 
 from backend.config import get_settings
 from backend.errors import ModelExecutionError
@@ -39,6 +43,8 @@ class MedASRModel:
         self.settings = get_settings()
         self.model_manager = model_manager or ModelManager()
         self._pipeline = None
+        self._processor = None
+        self._device = "cpu"
 
     @property
     def is_mock_mode(self) -> bool:
@@ -53,13 +59,13 @@ class MedASRModel:
         return self.settings.MEDASR_MODEL_ID.lower() == "mock"
 
     def load_model(self) -> None:
-        """Load the MedASR transformer pipeline unless running in mock mode.
+        """Load the MedASR model and processor unless running in mock mode.
 
         Args:
             None: Uses settings for model id and device selection.
 
         Returns:
-            None: Caches loaded pipeline instance.
+            None: Caches loaded model and processor instances.
         """
         if self.is_mock_mode:
             self._pipeline = "mock"
@@ -69,7 +75,7 @@ class MedASRModel:
         if self._pipeline is not None:
             return
 
-        if pipeline is None:
+        if AutoModelForCTC is None:
             raise ModelExecutionError("transformers is required for non-mock MedASR mode")
 
         device = "cuda:0"
@@ -77,11 +83,12 @@ class MedASRModel:
             device = "cpu"
 
         try:
-            self._pipeline = pipeline(
-                "automatic-speech-recognition",
-                model=self.settings.MEDASR_MODEL_ID,
-                device=device,
-            )
+            self._processor = AutoProcessor.from_pretrained(self.settings.MEDASR_MODEL_ID)
+            model = AutoModelForCTC.from_pretrained(self.settings.MEDASR_MODEL_ID)
+            model = model.to(device)
+            model.eval()
+            self._device = device
+            self._pipeline = model  # store model here so is_mock_mode / None checks still work
         except Exception as exc:
             raise ModelExecutionError(f"Failed to load MedASR model: {exc}") from exc
 
@@ -108,21 +115,66 @@ class MedASRModel:
             duration_s = self._duration(source)
             return self._make_transcript(source, text, duration_s)
 
-        waveform, _ = librosa.load(source, sr=16000, mono=True)
-        duration_s = float(librosa.get_duration(y=waveform, sr=16000))
+        waveform, file_sr = sf.read(source, dtype="float32", always_2d=False)
+        # Convert to mono if stereo
+        if waveform.ndim > 1:
+            waveform = waveform.mean(axis=1)
+        # Resample if needed
+        if file_sr != 16000:
+            from scipy.signal import resample
+
+            num_samples = int(len(waveform) * 16000 / file_sr)
+            waveform = resample(waveform, num_samples).astype(np.float32)
+        duration_s = float(len(waveform)) / 16000.0
 
         try:
-            result = self._pipeline(
+            inputs = self._processor(
                 waveform,
-                chunk_length_s=20,
-                stride_length_s=(4, 2),
-                return_timestamps=True,
-                generate_kwargs={"language": "en", "task": "transcribe"},
+                sampling_rate=16000,
+                return_tensors="pt",
+                padding=True,
             )
+            # MedASR processor may return input_features or input_values
+            if hasattr(inputs, "input_features") and inputs.input_features is not None:
+                model_input = inputs.input_features.to(self._device)
+            elif hasattr(inputs, "input_values") and inputs.input_values is not None:
+                model_input = inputs.input_values.to(self._device)
+            else:
+                # Fallback: get first tensor from the batch encoding
+                key = list(inputs.data.keys())[0]
+                model_input = inputs[key].to(self._device)
+
+            with torch.no_grad():
+                model_input = model_input.float()
+                logits = self._pipeline(**{list(inputs.data.keys())[0]: model_input}).logits
+
+            predicted_ids = torch.argmax(logits, dim=-1)
+
+            # CTC decoding: collapse consecutive duplicate tokens, then remove blanks
+            ids = predicted_ids[0].tolist()
+            collapsed = []
+            prev = None
+            for t in ids:
+                if t != prev:
+                    collapsed.append(t)
+                prev = t
+            # Token 0 is the CTC blank in most CTC models
+            blank_id = getattr(self._pipeline.config, 'ctc_blank_id', 0)
+            collapsed = [t for t in collapsed if t != blank_id]
+            collapsed_tensor = torch.tensor([collapsed], dtype=predicted_ids.dtype)
+
+            raw_text = self._processor.batch_decode(collapsed_tensor)[0]
+            # Strip any remaining special tokens
+            transcript_text = raw_text.replace("<epsilon>", "").replace("</s>", "").replace("<s>", "").strip()
+            # Collapse multiple spaces
+            import re as _re
+
+            transcript_text = _re.sub(r'\s+', ' ', transcript_text)
         except Exception as exc:
-            raise ModelExecutionError(f"MedASR inference failed: {exc}") from exc
+            import traceback
 
-        transcript_text = str(result.get("text", "")).strip()
+            tb = traceback.format_exc()
+            raise ModelExecutionError(f"MedASR inference failed: {exc}\nTraceback:\n{tb}") from exc
         return self._make_transcript(source, transcript_text, duration_s)
 
     def _make_transcript(self, audio_path: Path, text: str, duration_s: float) -> Transcript:
@@ -148,7 +200,7 @@ class MedASRModel:
 
     @staticmethod
     def _duration(audio_path: Path) -> float:
-        """Compute audio duration in seconds using librosa.
+        """Compute audio duration in seconds using soundfile.
 
         Args:
             audio_path (Path): Audio file path.
@@ -156,8 +208,8 @@ class MedASRModel:
         Returns:
             float: Duration in seconds.
         """
-        waveform, sample_rate = librosa.load(audio_path, sr=16000, mono=True)
-        return float(librosa.get_duration(y=waveform, sr=sample_rate))
+        info = sf.info(audio_path)
+        return float(info.frames) / float(info.samplerate)
 
     @staticmethod
     def _get_mock_text(audio_path: Path) -> str:

backend/orchestrator.py

@@ -180,7 +180,11 @@ class PipelineOrchestrator:
             raise ModelExecutionError("Audio could not be transcribed.")
         consultation.transcript = transcript.model_copy(update={"consultation_id": consultation_id})
         transcribe_s = round(time.perf_counter() - stage_start, 3)
+        if consultation.transcript and consultation.transcript.text:
+            logger.info("MedASR transcript:\n{}", consultation.transcript.text)
         logger.info("Pipeline stage complete", consultation_id=consultation_id, stage="transcribe", duration_s=transcribe_s)
+        if torch is not None and torch.cuda.is_available():
+            torch.cuda.empty_cache()
         self._clear_cuda_cache()
 
         stage_start = time.perf_counter()
@@ -198,7 +202,11 @@ class PipelineOrchestrator:
                 logger.warning("FHIR degradation activated", consultation_id=consultation_id, warning=warning)
                 consultation.context = self._build_transcript_only_context(consultation, warning)
         context_s = round(time.perf_counter() - stage_start, 3)
+        if consultation.context:
+            logger.info("EHR context extracted:\n{}", consultation.context.model_dump_json(indent=2)[:3000])
         logger.info("Pipeline stage complete", consultation_id=consultation_id, stage="retrieve_context", duration_s=context_s)
+        if torch is not None and torch.cuda.is_available():
+            torch.cuda.empty_cache()
         self._clear_cuda_cache()
 
         stage_start = time.perf_counter()
@@ -214,6 +222,8 @@ class PipelineOrchestrator:
                 consultation.transcript.text,
                 consultation.context,
                 consultation_id,
+                doc_type=consultation.doc_type,
+                letter_prefs=consultation.letter_prefs,
             ).model_copy(update={"consultation_id": consultation_id})
         else:
             raise ModelExecutionError("Transcript and patient context are required before document generation")
@@ -245,6 +255,8 @@ class PipelineOrchestrator:
         transcript_text: str,
         context: PatientContext,
         consultation_id: str,
+        doc_type: str = "Clinic Letter",
+        letter_prefs: dict | None = None,
     ) -> ClinicalDocument:
         """Generate a document with one OOM recovery retry.
 
@@ -259,7 +271,7 @@ class PipelineOrchestrator:
 
         max_tokens = int(self._doc_generator.settings.DOC_GEN_MAX_TOKENS)
         try:
-            return self._doc_generator.generate_document(transcript_text, context, max_new_tokens=max_tokens)
+            return self._doc_generator.generate_document(transcript_text, context, max_new_tokens=max_tokens, doc_type=doc_type, letter_prefs=letter_prefs)
         except TypeError as exc:
             if "max_new_tokens" not in str(exc):
                 raise
@@ -267,7 +279,7 @@ class PipelineOrchestrator:
                 "Document generator does not accept max_new_tokens override; falling back to default signature",
                 consultation_id=consultation_id,
             )
-            return self._doc_generator.generate_document(transcript_text, context)
+            return self._doc_generator.generate_document(transcript_text, context, doc_type=doc_type, letter_prefs=letter_prefs)
         except torch.cuda.OutOfMemoryError as exc:
             self._clear_cuda_cache()
             reduced_tokens = max(256, max_tokens // 2)
@@ -277,7 +289,7 @@ class PipelineOrchestrator:
                 previous_max_new_tokens=max_tokens,
                 retry_max_new_tokens=reduced_tokens,
             )
-            return self._doc_generator.generate_document(transcript_text, context, max_new_tokens=reduced_tokens)
+            return self._doc_generator.generate_document(transcript_text, context, max_new_tokens=reduced_tokens, doc_type=doc_type, letter_prefs=letter_prefs)
 
     def _build_transcript_only_context(self, consultation: Consultation, warning: str) -> PatientContext:
         """Build minimal patient context when EHR retrieval fails.

backend/prompts/document_generation.j2

@@ -3,9 +3,9 @@ You are an NHS clinical documentation assistant. Generate a structured NHS outpa
 
 STRICT OUTPUT FORMAT (follow exactly)
 - Date: {{ letter_date }}
-- Addressee: GP name and practice address from the patient record
-- Re: Full patient name, DOB, NHS number
-- Salutation: Dear Dr [GP surname],
+- Addressee: {{ gp_name }}, {{ gp_address }}
+- Re: {{ patient_name }}, DOB {{ patient_dob }}, NHS No. {{ patient_nhs }}
+- Salutation: Dear {{ gp_name }},
 - Body sections in this exact order:
   1) History of presenting complaint
   2) Examination findings
@@ -22,7 +22,7 @@ STYLE AND SAFETY RULES
 3. Keep length between 300 and 500 words.
 4. Use ONLY facts from transcript + patient context.
 5. Use EXACT numeric values from patient context (no rounding, no unit changes, no fabrication).
-6. If the transcript states a value that differs from EHR context, include the EHR value and mark the mismatch as [DISCREPANCY].
+6. If the transcript states a value that differs from EHR context, present BOTH values clearly. Use this exact format: "HbA1c of 8.2% (note: EHR recorded 7.8% on 01/11/2025)". Do NOT use square brackets, tags, or annotations like [DISCREPANCY].
 7. Do not add bullet points unless the source explicitly lists items.
 8. If a section has no discussed information, write a short factual sentence stating this.
 

backend/prompts/ward_round_generation.j2

@@ -0,0 +1,45 @@
+<|system|>
+You are an NHS clinical documentation assistant. Generate a structured ward round progress note from the consultation transcript and patient context provided below.
+
+STRICT OUTPUT FORMAT (follow exactly)
+- Date: {{ letter_date }}
+- Patient: {{ patient_name }}, DOB {{ patient_dob }}, NHS No. {{ patient_nhs }}
+- Ward / Bed: [as mentioned in transcript or "Not specified"]
+- Body sections in this exact order:
+  1) Overnight events
+  2) Current status and observations
+  3) Examination findings
+  4) Investigation results
+  5) Assessment and plan
+  6) Current medications
+  7) Tasks / Actions
+- Sign-off:
+  {{ clinician_name }}
+  {{ clinician_title }}
+
+STYLE AND SAFETY RULES
+1. Use formal British medical English, third person, past tense.
+2. Include both positive and negative findings from the consultation.
+3. Keep length between 200 and 400 words.
+4. Use ONLY facts from transcript + patient context.
+5. Use EXACT numeric values from patient context (no rounding, no unit changes, no fabrication).
+6. If a section has no discussed information, write "Not discussed."
+7. Do not use bullet points unless explicitly listing tasks.
+
+NEGATIVE EXAMPLES (DO NOT DO THESE)
+- Do NOT invent overnight events that were not discussed.
+- Do NOT replace exact values with vague language.
+- Do NOT use US spelling.
+<|end|>
+
+<|user|>
+## WARD ROUND TRANSCRIPT
+{{ transcript }}
+
+## PATIENT CONTEXT (from Electronic Health Record)
+{{ context_json }}
+
+Generate the ward round progress note now.
+<|end|>
+
+<|assistant|>

backend/schemas.py

@@ -128,6 +128,8 @@ class Consultation(BaseModel):
     started_at: Optional[str] = None
     ended_at: Optional[str] = None
     audio_file_path: Optional[str] = None
+    doc_type: str = Field(default="Clinic Letter", description="Document type: 'Clinic Letter' or 'Ward Round Note'")
+    letter_prefs: dict = Field(default_factory=dict, description="Letter preferences from frontend (clinician name, GP, etc.)")
 
 
 class PipelineProgress(BaseModel):

frontend/components.py

@@ -42,6 +42,8 @@ def build_global_style_block() -> str:
     return """
 <style>
   @import url('https://fonts.googleapis.com/css2?family=DM+Serif+Display:ital@0;1&family=Inter:wght@300;400;500;600;700&family=JetBrains+Mono:wght@400;500&display=swap');
+</style>
+<style>
   @keyframes clarkeGradientShift {
     0% { background-position: 0% 50%; }
     50% { background-position: 100% 50%; }
@@ -60,6 +62,195 @@ def build_global_style_block() -> str:
 
   html, body { margin: 0 !important; padding: 0 !important; overflow-x: hidden !important; }
 
+  /* Letter Preferences Accordion */
+  #clarke-letter-prefs {
+    margin: 0 48px 24px 48px !important;
+    border: 1px solid rgba(212, 175, 55, 0.25) !important;
+    border-radius: 12px !important;
+    background: rgba(255, 255, 255, 0.65) !important;
+    backdrop-filter: blur(8px) !important;
+    box-shadow: 0 2px 8px rgba(0, 0, 0, 0.03) !important;
+  }
+  #clarke-letter-prefs > .label-wrap {
+    padding: 14px 20px !important;
+    background: transparent !important;
+    border: none !important;
+    border-bottom: none !important;
+    cursor: pointer !important;
+  }
+  #clarke-letter-prefs > .label-wrap span {
+    font-family: 'DM Serif Display', serif !important;
+    font-size: 16px !important;
+    color: #D4AF37 !important;
+  }
+  #clarke-letter-prefs > .label-wrap:hover {
+    background: rgba(212, 175, 55, 0.04) !important;
+  }
+  #clarke-letter-prefs input[type="text"],
+  #clarke-letter-prefs textarea {
+    font-family: 'Inter', sans-serif !important;
+    font-size: 14px !important;
+    border: 1px solid rgba(212, 175, 55, 0.2) !important;
+    border-radius: 8px !important;
+    background: rgba(255, 255, 255, 0.85) !important;
+    color: #1A1A2E !important;
+    padding: 10px 14px !important;
+    transition: border-color 0.3s ease, box-shadow 0.3s ease !important;
+  }
+  #clarke-letter-prefs input[type="text"]:focus,
+  #clarke-letter-prefs textarea:focus {
+    border-color: #D4AF37 !important;
+    box-shadow: 0 0 0 3px rgba(212, 175, 55, 0.12) !important;
+    outline: none !important;
+  }
+  #clarke-letter-prefs label span {
+    font-family: 'Inter', sans-serif !important;
+    font-size: 13px !important;
+    font-weight: 600 !important;
+    color: #555 !important;
+  }
+
+  /* Letter Preferences Accordion */
+  #clarke-letter-prefs {
+    margin: 0 48px 24px 48px !important;
+    border: 1px solid rgba(212, 175, 55, 0.2) !important;
+    border-radius: 12px !important;
+    background: rgba(255, 255, 255, 0.6) !important;
+    backdrop-filter: blur(8px) !important;
+    overflow: hidden !important;
+  }
+  #clarke-letter-prefs > .label-wrap {
+    padding: 14px 20px !important;
+    background: transparent !important;
+    border: none !important;
+    cursor: pointer !important;
+  }
+  #clarke-letter-prefs > .label-wrap > span {
+    font-family: 'DM Serif Display', serif !important;
+    font-size: 16px !important;
+    color: #D4AF37 !important;
+  }
+  #clarke-letter-prefs > .label-wrap:hover {
+    background: rgba(212, 175, 55, 0.04) !important;
+  }
+  #clarke-letter-prefs .wrap {
+    padding: 4px 20px 16px 20px !important;
+    border-top: 1px solid rgba(212, 175, 55, 0.12) !important;
+  }
+  #clarke-letter-prefs input[type="text"],
+  #clarke-letter-prefs textarea {
+    font-family: 'Inter', sans-serif !important;
+    font-size: 14px !important;
+    border: 1px solid rgba(212, 175, 55, 0.2) !important;
+    border-radius: 8px !important;
+    background: rgba(255, 255, 255, 0.8) !important;
+    color: #1A1A2E !important;
+    padding: 10px 14px !important;
+    transition: all 0.3s ease !important;
+  }
+  #clarke-letter-prefs input[type="text"]:focus,
+  #clarke-letter-prefs textarea:focus {
+    border-color: #D4AF37 !important;
+    box-shadow: 0 0 0 2px rgba(212, 175, 55, 0.15) !important;
+    outline: none !important;
+  }
+  #clarke-letter-prefs label span {
+    font-family: 'Inter', sans-serif !important;
+    font-size: 13px !important;
+    font-weight: 600 !important;
+    color: #555 !important;
+  }
+
+  /* Document Type Radio — matches Letter Preferences */
+  fieldset#clarke-doc-type,
+  #clarke-doc-type {
+    --block-border-width: 0px !important;
+    --block-border-color: transparent !important;
+    --border-color-primary: transparent !important;
+    --block-background-fill: transparent !important;
+    --block-shadow: none !important;
+    --block-radius: 12px !important;
+    margin: 0 48px 16px 48px !important;
+    border: 1px solid rgba(212, 175, 55, 0.2) !important;
+    border-radius: 12px !important;
+    background: rgba(255, 255, 255, 0.6) !important;
+    backdrop-filter: blur(8px) !important;
+    -webkit-backdrop-filter: blur(8px) !important;
+    box-shadow: 0 2px 8px rgba(0, 0, 0, 0.03) !important;
+    padding: 0 !important;
+    overflow: hidden !important;
+  }
+  fieldset#clarke-doc-type > div,
+  #clarke-doc-type > div,
+  #clarke-doc-type .form {
+    background: transparent !important;
+    border: none !important;
+    box-shadow: none !important;
+    border-radius: 0 !important;
+  }
+  #clarke-doc-type > span[data-testid="block-info"] {
+    font-family: 'DM Serif Display', serif !important;
+    font-size: 16px !important;
+    color: #D4AF37 !important;
+    padding: 14px 20px 8px 20px !important;
+    display: block !important;
+  }
+  #clarke-doc-type > div.wrap {
+    padding: 4px 20px 16px 20px !important;
+    border-top: 1px solid rgba(212, 175, 55, 0.12) !important;
+    background: transparent !important;
+    gap: 12px !important;
+    border-left: none !important;
+    border-right: none !important;
+    border-bottom: none !important;
+    box-shadow: none !important;
+  }
+  fieldset#clarke-doc-type label,
+  #clarke-doc-type label {
+    font-family: 'DM Serif Display', serif !important;
+    font-size: 14px !important;
+    border: 1px solid rgba(212, 175, 55, 0.25) !important;
+    border-radius: 8px !important;
+    padding: 10px 20px !important;
+    cursor: pointer !important;
+    transition: all 0.3s ease !important;
+    background: rgba(255, 255, 255, 0.6) !important;
+    color: #555 !important;
+  }
+  #clarke-doc-type label.selected {
+    background: rgba(212, 175, 55, 0.12) !important;
+    border-color: #D4AF37 !important;
+    color: #1A1A2E !important;
+    font-weight: 600 !important;
+  }
+  #clarke-doc-type label:hover {
+    background: rgba(212, 175, 55, 0.06) !important;
+    border-color: rgba(212, 175, 55, 0.4) !important;
+  }
+  #clarke-doc-type input[type="radio"] {
+    accent-color: #D4AF37 !important;
+  }
+  #clarke-doc-type *:not(label):not(input):not(span) {
+    background: transparent !important;
+    border-color: transparent !important;
+    box-shadow: none !important;
+  }
+  #clarke-doc-type > div[class*="form"],
+  #clarke-doc-type > div[class*="svelte"] {
+    background: transparent !important;
+    border: none !important;
+    box-shadow: none !important;
+    overflow: visible !important;
+  }
+
+  /* Make Gradio progress bar gold instead of red */
+  .progress-bar, .progress-bar > .progress-bar-wrap, .progress-bar > .progress-bar-wrap > .progress-bar-fill {
+    background: linear-gradient(135deg, #D4AF37, #F0D060) !important;
+  }
+  .eta-bar { background: rgba(212, 175, 55, 0.15) !important; }
+  /* Hide Gradio error toasts */
+  .toast-wrap, .toast-body, .error { display: none !important; }
+
   #hidden-select-0, #hidden-select-1, #hidden-select-2, #hidden-select-3, #hidden-select-4,
   #hidden-start-consultation, #hidden-back, #hidden-cancel, #hidden-regenerate, #hidden-copy, #hidden-download,
   #hidden-end-consultation, #hidden-sign-off, #hidden-next-patient {
@@ -148,7 +339,7 @@ def build_global_style_block() -> str:
     overflow: hidden !important;
   }
 </style>
-<img src="data:image/gif;base64,R0lGODlhAQABAIAAAAAAAP///yH5BAEAAAAALAAAAAABAAEAAAIBRAA7" onload="(function(){function e(){document.documentElement.style.setProperty('background','#F8F6F1','important');var a=document.querySelector('gradio-app');if(a){a.style.setProperty('background','transparent','important');a.style.setProperty('padding','0','important');a.style.setProperty('margin','0','important');a.style.setProperty('overflow-x','hidden','important');}document.querySelectorAll('.gradio-container,[class*=gradio-container-]').forEach(function(c){c.style.setProperty('max-width','100vw','important');c.style.setProperty('padding','0','important');c.style.setProperty('margin','0','important');c.style.setProperty('background','transparent','important');});document.body.style.setProperty('margin','0','important');document.body.style.setProperty('padding','0','important');document.body.style.setProperty('background','transparent','important');var f=document.querySelector('footer');if(f)f.style.display='none';}if(!document.getElementById('clarke-sunrise-glow')){var s=document.createElement('style');s.textContent='@keyframes clarkeWarmthPulse{0%{opacity:0.55;transform:scaleY(1) scaleX(1);}50%{opacity:1;transform:scaleY(1.35) scaleX(1.12);}100%{opacity:0.55;transform:scaleY(1) scaleX(1);}}';document.head.appendChild(s);var g=document.createElement('div');g.id='clarke-sunrise-glow';g.style.cssText='position:fixed;top:0;left:0;width:100vw;height:600px;pointer-events:none;z-index:0;background:radial-gradient(ellipse 140% 110% at 50% 0%, rgba(255,193,7,0.80) 0%, rgba(255,213,79,0.55) 20%, rgba(212,175,55,0.28) 45%, transparent 75%);animation:clarkeWarmthPulse 8s ease-in-out infinite;transform-origin:top center;';document.body.insertBefore(g,document.body.firstChild);console.log('Clarke: Sunrise glow injected');}e();[100,300,600,1200,2500,5000].forEach(function(t){setTimeout(e,t);});new MutationObserver(function(){e();}).observe(document.documentElement,{childList:true,subtree:true,attributes:true,attributeFilter:['style','class']});console.log('Clarke: Layout enforcer active via img onload');})()" style="display:none;position:absolute;width:0;height:0;">
+<img src="data:image/gif;base64,R0lGODlhAQABAIAAAAAAAP///yH5BAEAAAAALAAAAAABAAEAAAIBRAA7" onload="(function(){function e(){document.documentElement.style.setProperty('background','#F8F6F1','important');var a=document.querySelector('gradio-app');if(a){a.style.setProperty('background','transparent','important');a.style.setProperty('padding','0','important');a.style.setProperty('margin','0','important');a.style.setProperty('overflow-x','hidden','important');}document.querySelectorAll('.gradio-container,[class*=gradio-container-]').forEach(function(c){c.style.setProperty('max-width','100vw','important');c.style.setProperty('padding','0','important');c.style.setProperty('margin','0','important');c.style.setProperty('background','transparent','important');});document.body.style.setProperty('margin','0','important');document.body.style.setProperty('padding','0','important');document.body.style.setProperty('background','transparent','important');var f=document.querySelector('footer');if(f)f.style.display='none';}if(!document.getElementById('clarke-sunrise-glow')){var s=document.createElement('style');s.textContent='@keyframes clarkeWarmthPulse{0%{opacity:0.55;transform:scaleY(1) scaleX(1);}50%{opacity:1;transform:scaleY(1.35) scaleX(1.12);}100%{opacity:0.55;transform:scaleY(1) scaleX(1);}}';document.head.appendChild(s);var g=document.createElement('div');g.id='clarke-sunrise-glow';g.style.cssText='position:fixed;top:0;left:0;width:100vw;height:600px;pointer-events:none;z-index:0;background:radial-gradient(ellipse 140% 110% at 50% 0%, rgba(255,193,7,0.80) 0%, rgba(255,213,79,0.55) 20%, rgba(212,175,55,0.28) 45%, transparent 75%);animation:clarkeWarmthPulse 8s ease-in-out infinite;transform-origin:top center;';document.body.insertBefore(g,document.body.firstChild);console.log('Clarke: Sunrise glow injected');}window.clarkePrintPDF=function(){var el=document.getElementById('signed-letter-text');var text='';if(el){text=el.innerText||el.textContent;}if(!text){alert('No letter text found');return;}text=text.trim();var headings=['History of Presenting Complaint','Examination','Investigations','Assessment','Plan'];var lines=text.split('\\n');var css='@page{size:A4;margin:25mm 20mm 25mm 20mm;}body{font-family:Helvetica,Arial,sans-serif;font-size:12pt;line-height:1.6;color:#1a1a2e;margin:0;padding:0;}.hdr{border-top:3px solid #D4AF37;margin-bottom:8px;}.trust{text-align:right;color:#888;font-size:12pt;margin-bottom:16px;}.gl{border-top:1.5px solid #D4AF37;margin:12px 0;}.sh{font-weight:bold;font-size:14pt;color:#1a1a2e;margin-top:20px;margin-bottom:4px;border-bottom:2px solid #D4AF37;display:inline-block;padding-bottom:2px;}.rl{font-weight:bold;font-size:13pt;}.pi{margin-left:12px;}.so{margin-top:24px;}.sn{font-weight:bold;}.ft{margin-top:40px;text-align:center;color:#bbb;font-size:9pt;}';var h='<!DOCTYPE html><html><head><style>'+css+'</style></head><body>';h+='<div class=hdr></div>';h+='<div class=trust>Clarke NHS Trust<br>General Practice Department<br>University Hospital London</div>';h+='<div class=gl></div>';var inSignoff=false;for(var i=0;i<lines.length;i++){var line=lines[i].trim();if(!line){h+='<br>';continue;}var isH=false;for(var j=0;j<headings.length;j++){if(line===headings[j]){isH=true;break;}}if(isH){h+='<div class=sh>'+line+'</div>';continue;}if(line.match(/^Re:/)){h+='<div class=rl>'+line+'</div>';continue;}if(line.match(/^Warm regards/)||line.match(/^Yours sincerely/)){inSignoff=true;h+='<div class=so>'+line+'</div>';continue;}if(inSignoff){h+='<div class=sn>'+line+'</div>';continue;}if(line.match(/^\d+\./)){h+='<div class=pi>'+line+'</div>';continue;}h+='<div>'+line+'</div>';}h+='<div class=ft>Generated by Clarke - AI Clinical Documentation System</div></body></html>';var iframe=document.createElement('iframe');iframe.style.cssText='position:fixed;top:-9999px;left:-9999px;width:210mm;height:297mm;';document.body.appendChild(iframe);iframe.contentDocument.open();iframe.contentDocument.write(h);iframe.contentDocument.close();setTimeout(function(){iframe.contentWindow.print();setTimeout(function(){document.body.removeChild(iframe);},2000);},500);console.log('Clarke: Print PDF dialog opened');};console.log('Clarke: clarkePrintPDF registered');e();[100,300,600,1200,2500,5000].forEach(function(t){setTimeout(e,t);});new MutationObserver(function(){e();}).observe(document.documentElement,{childList:true,subtree:true,attributes:true,attributeFilter:['style','class']});console.log('Clarke: Layout enforcer active via img onload');})()" style="display:none;position:absolute;width:0;height:0;">
 """
 
 

frontend/state.py

@@ -34,6 +34,16 @@ def initial_consultation_state() -> dict[str, Any]:
         "current_patient_index": 0,
         "completed_patients": [],
         "signed_letters": {},
+        "doc_type": "Clinic Letter",
+        "letter_prefs": {
+            "clinician_name": "Dr Sarah Chen",
+            "clinician_title": "Consultant, General Practice",
+            "hospital": "Clarke NHS Trust",
+            "department": "General Practice Department",
+            "gp_name": "Dr Andrew Wilson",
+            "gp_address": "Riverside Medical Practice\n14 Harcourt Street\nLondon",
+            "signoff_phrase": "Warm regards",
+        },
     }
 
 

frontend/ui.py

@@ -351,6 +351,64 @@ def _format_patient_context_html(context: dict[str, Any]) -> str:
     )
 
 
+def _letter_prefs_persistence_js() -> str:
+    """Return an HTML snippet that persists letter preference values via JavaScript."""
+    return """<img src="data:image/gif;base64,R0lGODlhAQABAIAAAAAAAP///yH5BAEAAAAALAAAAAABAAEAAAIBRAA7" onload="(function(){
+        if(window.clarkePrefsInitDone)return;
+        window.clarkePrefsInitDone=true;
+        window.clarkeLetterPrefs={};
+        function getInputs(){
+            var acc=document.getElementById('clarke-letter-prefs');
+            if(!acc)return[];
+            return Array.prototype.slice.call(acc.querySelectorAll('input[type=text],textarea'));
+        }
+        function saveAll(){
+            var inputs=getInputs();
+            for(var i=0;i<inputs.length;i++){
+                window.clarkeLetterPrefs[i]=inputs[i].value;
+            }
+        }
+        function restoreAll(){
+            var inputs=getInputs();
+            if(inputs.length===0)return;
+            var changed=false;
+            for(var i=0;i<inputs.length;i++){
+                var saved=window.clarkeLetterPrefs[i];
+                if(saved!==undefined&&saved!==inputs[i].value){
+                    var proto=inputs[i].tagName==='TEXTAREA'?window.HTMLTextAreaElement.prototype:window.HTMLInputElement.prototype;
+                    var setter=Object.getOwnPropertyDescriptor(proto,'value');
+                    if(setter&&setter.set){setter.set.call(inputs[i],saved);}
+                    else{inputs[i].value=saved;}
+                    inputs[i].dispatchEvent(new Event('input',{bubbles:true}));
+                    inputs[i].dispatchEvent(new Event('change',{bubbles:true}));
+                    changed=true;
+                }
+            }
+            if(changed)console.log('Clarke: Restored letter prefs');
+        }
+        function attachListeners(){
+            var inputs=getInputs();
+            inputs.forEach(function(inp){
+                if(!inp.dataset.clarkeTracked){
+                    inp.dataset.clarkeTracked='1';
+                    inp.addEventListener('input',function(){saveAll();});
+                    inp.addEventListener('change',function(){saveAll();});
+                }
+            });
+        }
+        function check(){
+            var inputs=getInputs();
+            if(inputs.length>0){
+                attachListeners();
+                if(Object.keys(window.clarkeLetterPrefs).length>0){restoreAll();}
+            }
+        }
+        new MutationObserver(function(){check();}).observe(document.body,{childList:true,subtree:true});
+        setInterval(check,2000);
+        console.log('Clarke: Letter prefs persistence active');
+    })()" style="display:none;position:absolute;width:0;height:0;">"""
+
+
 def _context_screen_html(patient: dict[str, Any], context: dict[str, Any]) -> str:
     """Build S2 shell + actions + context in one full-screen HTML block."""
 
@@ -382,8 +440,13 @@ def _build_generated_document(state: dict[str, Any]) -> dict[str, Any]:
     dob = str(demographics.get("dob") or "Unknown")
     nhs = str(demographics.get("nhs_number") or "Unknown")
     today = datetime.now().strftime("%d %B %Y")
-    gp_name = "Andrew Wilson"
-    address = "Riverside Medical Practice\n14 Harcourt Street\nLondon"
+    prefs = (state or {}).get("letter_prefs", {})
+    gp_name = prefs.get("gp_name", "Dr Andrew Wilson").replace("Dr ", "", 1)
+    address = prefs.get("gp_address", "Riverside Medical Practice\n14 Harcourt Street\nLondon")
+    clinician_display = prefs.get("clinician_name", "Dr Sarah Chen")
+    clinician_title = prefs.get("clinician_title", "Consultant, General Practice")
+    hospital_name = prefs.get("hospital", "Clarke NHS Trust")
+    signoff_phrase = prefs.get("signoff_phrase", "Warm regards")
 
     investigations = "\n".join(
         f"- {lab.get('name', 'Test')}: {lab.get('value', '')} {lab.get('unit', '')} ({lab.get('date', '')})".strip()
@@ -409,6 +472,8 @@ def _build_generated_document(state: dict[str, Any]) -> dict[str, Any]:
             "Review in specialist clinic to reassess response and escalation needs.",
         ]
 
+    clinical_issues_text = "\n".join(f"- {item}" for item in problems) if problems else "- None listed"
+
     letter_text = (
         f"{today}\n\n"
         f"Dr {gp_name}\n"
@@ -416,29 +481,98 @@ def _build_generated_document(state: dict[str, Any]) -> dict[str, Any]:
         f"Dear Dr {gp_name},\n\n"
         f"Re: {patient_name} (DOB: {dob}, NHS: {nhs})\n"
         f"    {address}\n\n"
-        f"Thank you for referring / I reviewed {patient_name} in General Practice Clinic on {today}.\n\n"
+        f"Thank you for referring / I reviewed {patient_name} in {clinician_title.split(',')[-1].strip() if ',' in clinician_title else clinician_title} Clinic on {today}.\n\n"
         "History of Presenting Complaint\n"
         f"{history}\n\n"
         "Examination\n"
         "The patient was comfortable at rest, haemodynamically stable, and clinically euvolaemic on examination. No acute red-flag findings were identified today.\n\n"
         "Investigations\n"
         f"{investigations}\n\n"
+        "Clinical Issues\n"
+        f"{clinical_issues_text}\n\n"
+        "Current Medications\n"
+        f"{medication_line or 'None documented'}\n\n"
         "Assessment\n"
         f"{assessment}\n\n"
         "Plan\n"
         + "\n".join(f"{i + 1}. {line}" for i, line in enumerate(plan_lines))
         + f"\n\nI will review {patient_name} in 8 weeks. Please do not hesitate to contact us if there are any concerns in the interim.\n\n"
-        "Yours sincerely,\n\n"
-        "Dr Sarah Chen\n"
-        "Consultant, General Practice\n"
-        "Clarke NHS Trust"
+        f"{signoff_phrase},\n\n"
+        f"{clinician_display}\n"
+        f"{clinician_title}\n"
+        f"{hospital_name}"
     )
 
+    doc_type = (state or {}).get("doc_type", "Clinic Letter")
+
+    if doc_type == "Ward Round Note":
+        now_time = datetime.now().strftime("%H:%M")
+
+        if "Margaret Thompson" in patient_name:
+            overnight = "Remained stable overnight. Blood glucose levels ranged 8.4-14.2 mmol/L. No hypoglycaemic episodes. Nursing staff report adequate oral intake."
+            current_status = "Alert and oriented. Reports mild fatigue but no chest pain, dyspnoea, or new symptoms. Tolerating diet well."
+            exam_findings = "Obs: BP 142/88, HR 78 regular, SpO2 97% RA, Temp 36.8. CVS: HS I+II+0, no peripheral oedema. Resp: Clear bilaterally. Abdo: Soft, non-tender."
+            today_plan = [
+                "Optimise glycaemic control — consider increasing gliclazide to 80mg BD.",
+                "Chase repeat HbA1c and renal profile results from this morning.",
+                "Dietitian review requested for structured carbohydrate counselling.",
+                "Continue current medications including lisinopril 10mg OD and atorvastatin 40mg ON.",
+                "Aim for discharge tomorrow if glucose control improving — arrange diabetes nurse follow-up within 1 week.",
+            ]
+        else:
+            overnight = f"Stable overnight. No acute events reported by nursing staff. Observations within acceptable parameters for {main_problem.lower()}."
+            current_status = f"Patient reports feeling stable this morning. Ongoing management of {main_problem.lower()} continues."
+            exam_findings = "Obs: Within normal limits. Systems examination unremarkable. No new clinical findings."
+            today_plan = [
+                "Continue current management plan.",
+                "Review outstanding investigation results.",
+                "Reassess clinical progress and escalation needs.",
+                "Estimated discharge: pending clinical improvement.",
+            ]
+
+        clinical_issues_ward = "\n".join(f"- {item}" for item in problems) if problems else "- None listed"
+
+        ward_note_text = (
+            f"WARD ROUND NOTE — {today} at {now_time}\n"
+            f"{'=' * 50}\n\n"
+            f"Patient: {patient_name}\n"
+            f"DOB: {dob} | NHS: {nhs}\n"
+            f"Ward: General Medical | Bed: 12A\n"
+            f"Consultant: {clinician_display}\n\n"
+            f"Day {2} of admission | Primary Dx: {main_problem}\n\n"
+            "Overnight Events\n"
+            f"{overnight}\n\n"
+            "Current Status\n"
+            f"{current_status}\n\n"
+            "Examination Findings\n"
+            f"{exam_findings}\n\n"
+            "Investigations\n"
+            f"{investigations}\n\n"
+            "Current Medications\n"
+            f"{medication_line or 'As per drug chart'}\n\n"
+            "Clinical Issues\n"
+            f"{clinical_issues_ward}\n\n"
+            "Assessment\n"
+            f"{assessment}\n\n"
+            "Plan\n"
+            + "\n".join(f"{i + 1}. {line}" for i, line in enumerate(today_plan))
+            + f"\n\n{clinician_display} | {clinician_title} | {hospital_name}\n"
+            f"Documented at {now_time} on {today}"
+        )
+
+        sections = [
+            {"heading": "Ward Round Note", "content": ward_note_text},
+        ]
+        return {
+            "title": "Ward Round Note",
+            "status": "ready_for_review",
+            "sections": sections,
+            "patient_name": patient_name,
+            "nhs_number": nhs,
+        }
+
     sections = [
         {"heading": "NHS Clinic Letter", "content": letter_text},
-        {"heading": "Clinical Issues", "content": "\n".join(f"- {item}" for item in problems) or "- None listed"},
-        {"heading": "Current Medications", "content": medication_line or "None documented"},
-        {"heading": "Follow-up", "content": "Review in 8 weeks with repeat investigations."},
     ]
     return {
         "title": "NHS Clinic Letter",
@@ -470,7 +604,7 @@ def _render_letter_sections(letter_sections: list[dict[str, str]]) -> tuple[str,
     return (combined, "", "", "")
 
 
-def _handle_patient_selection(state: dict[str, Any], patient_index: int):
+def _handle_patient_selection(state: dict[str, Any], patient_index: int, clinician_name: str = "Dr Sarah Chen", clinician_title: str = "Consultant, General Practice", hospital: str = "Clarke NHS Trust", department: str = "General Practice Department", gp_name: str = "Dr Andrew Wilson", signoff_phrase: str = "Warm regards", gp_address: str = "Riverside Medical Practice\n14 Harcourt Street\nLondon", doc_type: str = "Clinic Letter"):
     """Update state and call backend context endpoint when a patient index is selected.
 
     Args:
@@ -489,6 +623,16 @@ def _handle_patient_selection(state: dict[str, Any], patient_index: int):
     patient = patients[patient_index]
     updated_state = select_patient(state, patient)
     updated_state['current_patient_index'] = patient_index
+    updated_state["letter_prefs"] = {
+        "clinician_name": clinician_name or "Dr Sarah Chen",
+        "clinician_title": clinician_title or "Consultant, General Practice",
+        "hospital": hospital or "Clarke NHS Trust",
+        "department": department or "General Practice Department",
+        "gp_name": gp_name or "Dr Andrew Wilson",
+        "gp_address": gp_address or "Riverside Medical Practice\n14 Harcourt Street\nLondon",
+        "signoff_phrase": signoff_phrase or "Warm regards",
+    }
+    updated_state["doc_type"] = doc_type or "Clinic Letter"
     patient_id = str(patient.get("id", ""))
     if os.getenv("USE_MOCK_FHIR", "").lower() == "true":
         context = _mock_context_for_index(patient_index)
@@ -587,23 +731,37 @@ def _stage_from_pipeline(stage: str) -> tuple[int, str, str]:
     mapping = {
         "transcribing": (1, "Finalising transcript…", "MedASR processing audio"),
         "retrieving_context": (2, "Synthesising patient context…", "MedGemma 4B querying records"),
-        "generating_document": (3, "Generating clinical letter…", "MedGemma 27B composing document"),
-        "complete": (3, "Generating clinical letter…", "MedGemma 27B composing document"),
+        "generating_document": (3, "Generating document…", "MedGemma 27B composing document"),
+        "complete": (3, "Generating document…", "MedGemma 27B composing document"),
     }
     return mapping.get(stage, mapping["transcribing"])
 
 
 
 
-def _ensure_mock_audio_file(audio_path: str | None) -> str | None:
-    """Create a short silent WAV when running in mock mode and no audio was captured."""
+def _ensure_mock_audio_file(audio_path: str | None, state: dict | None = None) -> str | None:
+    """Return audio path, falling back to demo audio files for known patients."""
 
     if audio_path:
         return audio_path
-    if os.getenv("MEDASR_MODEL_ID", "").lower() != "mock":
-        return None
 
-    upload_dir = Path("data/uploads/mock")
+    # Map patient indices to demo audio files
+    DEMO_AUDIO_MAP = {
+        0: "data/demo/mrs_thompson.wav",
+        1: "data/demo/mr_okafor.wav",
+        2: "data/demo/ms_patel.wav",
+        3: "data/demo/mr_williams.wav",
+        4: "data/demo/mrs_khan.wav",
+    }
+
+    patient_index = (state or {}).get("current_patient_index")
+    if patient_index is not None and patient_index in DEMO_AUDIO_MAP:
+        demo_path = Path(DEMO_AUDIO_MAP[patient_index])
+        if demo_path.exists():
+            return str(demo_path)
+
+    # Fallback: generate a short silent WAV for patients without demo audio
+    upload_dir = Path("/tmp/mock_audio")
     upload_dir.mkdir(parents=True, exist_ok=True)
     silent_path = upload_dir / "silent.wav"
     with wave.open(str(silent_path), "wb") as wav_file:
@@ -631,7 +789,7 @@ def _start_processing(state, audio_path):
     if not consultation_id:
         return updated_state, "Consultation session is missing. Start consultation again.", _processing_screen_html(1, "Finalising transcript…", "MedASR processing audio", "Elapsed: 00:00"), gr.update(active=False), *show_screen("s3")
 
-    resolved_audio_path = _ensure_mock_audio_file(audio_path)
+    resolved_audio_path = _ensure_mock_audio_file(audio_path, state=updated_state)
     if not resolved_audio_path:
         return updated_state, "Please capture audio before ending consultation.", _processing_screen_html(1, "Finalising transcript…", "MedASR processing audio", "Elapsed: 00:00"), gr.update(active=False), *show_screen("s3")
 
@@ -645,9 +803,21 @@ def _start_processing(state, audio_path):
         return updated_state, "Consultation ended. Processing audio and generating document.", _processing_screen_html(1, "Finalising transcript…", "MedASR processing audio", "Elapsed: 00:00"), gr.update(active=True), *show_screen("s4")
 
     try:
-        with Path(resolved_audio_path).open("rb") as stream:
-            _api_request("POST", f"/consultations/{consultation_id}/audio", files={"audio_file": (Path(resolved_audio_path).name, stream, "audio/wav")}, data={"is_final": "true"}, timeout=120.0)
-        _api_request("POST", f"/consultations/{consultation_id}/end", timeout=180.0)
+        _api_request(
+            "POST",
+            f"/consultations/{consultation_id}/end",
+            json={
+                "audio_path": resolved_audio_path,
+                "doc_type": updated_state.get("doc_type", "Clinic Letter"),
+                "letter_prefs": {
+                    "clinician_name": updated_state.get("clinician_name", "Dr Sarah Chen"),
+                    "clinician_title": updated_state.get("clinician_title", "Consultant, General Practice"),
+                    "gp_name": updated_state.get("gp_name", "Dr Andrew Wilson"),
+                    "gp_address": updated_state.get("gp_address", "Riverside Medical Practice"),
+                },
+            },
+            timeout=300.0,
+        )
     except Exception as exc:
         return updated_state, f"Failed to end consultation: {exc}", _processing_screen_html(1, "Finalising transcript…", "MedASR processing audio", "Elapsed: 00:00"), gr.update(active=False), *show_screen("s3")
 
@@ -685,7 +855,7 @@ def _poll_processing_progress(state):
         updated_state["screen"] = "s5"
         s1, s2, s3, s4 = _render_letter_sections(doc.get("sections", []))
         fhir = ""
-        return updated_state, "Processing complete. Review the generated clinic letter.", _processing_screen_html(3, "Generating clinical letter…", "MedGemma 27B composing document", elapsed), gr.update(active=False), s1, s2, s3, s4, fhir, *show_screen("s5")
+        return updated_state, "Processing complete. Review the generated clinic letter.", _processing_screen_html(3, f"Generating {updated_state.get('doc_type', 'clinical letter').lower()}...", "MedGemma 27B composing document", elapsed), gr.update(active=False), s1, s2, s3, s4, fhir, *show_screen("s5")
 
     try:
         progress = _api_request("GET", f"/consultations/{consultation_id}/progress")
@@ -707,7 +877,7 @@ def _poll_processing_progress(state):
             f"<span style='font-family:JetBrains Mono,monospace;font-size:14px;background:rgba(212,175,55,0.1);padding:2px 6px;border-radius:4px;color:#1E3A8A;'>Patient: {escape(str(document_payload.get('patient_name', 'N/A')))}</span>",
         ]
     )
-    return updated_state, "Processing complete. Review the generated clinic letter.", _processing_screen_html(3, "Generating clinical letter…", "MedGemma 27B composing document", elapsed), gr.update(active=False), s1, s2, s3, s4, fhir, *show_screen("s5")
+    return updated_state, "Processing complete. Review the generated clinic letter.", _processing_screen_html(3, f"Generating {updated_state.get('doc_type', 'clinical letter').lower()}...", "MedGemma 27B composing document", elapsed), gr.update(active=False), s1, s2, s3, s4, fhir, *show_screen("s5")
 
 
 def _regenerate_document(state):
@@ -787,7 +957,7 @@ def _sign_off_document(state, section_1, section_2, section_3, section_4):
         updated_state["consultation"] = {"id": None, "status": "idle"}
     updated_state["consultation"]["status"] = "signed_off"
     updated_state["screen"] = "s6"
-    export_path = Path("data") / "demo" / "latest_signed_letter.txt"
+    export_path = Path("/tmp") / "latest_signed_letter.txt"
     export_path.write_text(signed_letter + "\n", encoding="utf-8")
     signed_html = f"<div style='min-height:100vh;background:#F8F6F1;padding:24px 48px 48px 48px;margin:0;'><div style='font-family:Inter,sans-serif;font-size:16px;line-height:1.75;color:#1A1A2E;white-space:pre-wrap;' id='signed-letter-text'>{escape(signed_letter)}</div></div>"
     return updated_state, "Document signed off. You can now copy or download the letter.", signed_html, signed_letter, gr.update(value=str(export_path)), *show_screen("s6")
@@ -826,11 +996,11 @@ def _prepare_signed_download(state):
     if not signed_text:
         return updated_state, "No signed letter available to download yet.", gr.update(value=None)
 
-    export_path = Path("data") / "demo" / "latest_signed_letter.txt"
+    export_path = Path("/tmp") / "latest_signed_letter.txt"
     export_path.write_text(signed_text + "\n", encoding="utf-8")
     return updated_state, "Download file refreshed.", gr.update(value=str(export_path))
 
-def _next_patient(state):
+def _next_patient(state, p_cn, p_ct, p_ho, p_de, p_gp, p_so, p_ga, p_dt):
     """Reset consultation workflow and return to dashboard after sign-off.
 
     Args:
@@ -850,7 +1020,35 @@ def _next_patient(state):
     refreshed_state = initial_consultation_state()
     refreshed_state['completed_patients'] = updated_state['completed_patients']
     refreshed_state['signed_letters'] = dict(updated_state.get('signed_letters', {}))
-    return refreshed_state, "Ready for next patient. Please select a patient card.", "", "", "", "", "", "", dashboard, *show_screen("s1")
+    refreshed_state['doc_type'] = p_dt or updated_state.get('doc_type', 'Clinic Letter')
+    refreshed_state['letter_prefs'] = {
+        "clinician_name": p_cn or "Dr Sarah Chen",
+        "clinician_title": p_ct or "Consultant, General Practice",
+        "hospital": p_ho or "Clarke NHS Trust",
+        "department": p_de or "General Practice Department",
+        "gp_name": p_gp or "Dr Andrew Wilson",
+        "signoff_phrase": p_so or "Warm regards",
+        "gp_address": p_ga or "Riverside Medical Practice\n14 Harcourt Street\nLondon",
+    }
+    return (
+        refreshed_state,
+        "Ready for next patient. Please select a patient card.",
+        "",
+        "",
+        "",
+        "",
+        "",
+        "",
+        dashboard,
+        *show_screen("s1"),
+        gr.update(),
+        gr.update(),
+        gr.update(),
+        gr.update(),
+        gr.update(),
+        gr.update(),
+        gr.update(),
+    )
 
 
 def build_ui() -> gr.Blocks:
@@ -868,6 +1066,7 @@ def build_ui() -> gr.Blocks:
     with gr.Blocks(theme=clarke_theme, css=Path("frontend/assets/style.css").read_text(encoding="utf-8"), title="Clarke", head=CLARKE_HEAD) as demo:
         app_state = gr.State(initial_consultation_state())
         gr.HTML(build_global_style_block())
+        gr.HTML("""<img src="data:image/gif;base64,R0lGODlhAQABAIAAAAAAAP///yH5BAEAAAAALAAAAAABAAEAAAIBRAA7" onload="(function(){function fix(){var el=document.getElementById('clarke-doc-type');if(!el)return;var p=el.parentElement;while(p&&p!==document.body){p.style.setProperty('background','transparent','important');p.style.setProperty('border','none','important');p.style.setProperty('box-shadow','none','important');p.style.setProperty('padding','0','important');if(p.classList.contains('form')||p.classList.contains('block'))break;p=p.parentElement;}el.style.setProperty('background','rgba(255,255,255,0.6)','important');el.style.setProperty('border','1px solid rgba(212,175,55,0.2)','important');el.style.setProperty('border-radius','12px','important');el.style.setProperty('backdrop-filter','blur(8px)','important');el.style.setProperty('-webkit-backdrop-filter','blur(8px)','important');el.style.setProperty('box-shadow','0 2px 8px rgba(0,0,0,0.03)','important');el.style.setProperty('overflow','hidden','important');el.style.setProperty('padding','0','important');var kids=el.querySelectorAll('div,fieldset');for(var i=0;i<kids.length;i++){kids[i].style.setProperty('background','transparent','important');kids[i].style.setProperty('border-color','transparent','important');kids[i].style.setProperty('box-shadow','none','important');}var wrap=el.querySelector('.wrap');if(wrap){wrap.style.setProperty('padding','4px 20px 16px 20px','important');wrap.style.setProperty('border-top','1px solid rgba(212,175,55,0.12)','important');wrap.style.setProperty('gap','12px','important');wrap.style.setProperty('background','transparent','important');}var title=el.querySelector('span[data-testid=block-info]');if(title){title.style.setProperty('font-family','DM Serif Display,serif','important');title.style.setProperty('font-size','16px','important');title.style.setProperty('color','#D4AF37','important');title.style.setProperty('padding','14px 20px 8px 20px','important');title.style.setProperty('display','block','important');}var labels=el.querySelectorAll('label');for(var j=0;j<labels.length;j++){labels[j].style.setProperty('font-family','DM Serif Display,serif','important');labels[j].style.setProperty('font-size','14px','important');labels[j].style.setProperty('border','1px solid rgba(212,175,55,0.25)','important');labels[j].style.setProperty('border-radius','8px','important');labels[j].style.setProperty('padding','10px 20px','important');labels[j].style.setProperty('cursor','pointer','important');labels[j].style.setProperty('background','rgba(255,255,255,0.6)','important');labels[j].style.setProperty('color','#555','important');}var radios=el.querySelectorAll('input[type=radio]');for(var k=0;k<radios.length;k++){radios[k].style.setProperty('accent-color','#D4AF37','important');}}fix();[100,300,600,1200,2500].forEach(function(t){setTimeout(fix,t);});new MutationObserver(function(){fix();}).observe(document.documentElement,{childList:true,subtree:true});})()" style="display:none;position:absolute;width:0;height:0;">""")
         feedback_text = gr.Markdown("", visible=False)
 
         with gr.Column(visible=False) as screen_s2:
@@ -888,10 +1087,10 @@ def build_ui() -> gr.Blocks:
             hidden_cancel_button = gr.Button("hidden-cancel", visible=True, elem_id="hidden-cancel")
 
         with gr.Column(visible=False) as screen_s5:
-            gr.HTML("<div style='min-height:100vh;background:#F8F6F1;padding:32px 48px;margin:0;'><h2 style='font-family:DM Serif Display,serif;color:#1A1A2E;margin:0 0 16px 0;'>Document Review</h2></div>")
+            gr.HTML("<h2 style='font-family:DM Serif Display,serif;color:#1A1A2E;margin:0;padding:8px 0 0 0;'>Document Review</h2>")
             review_status_badge = gr.HTML(build_status_badge_html("✎ Ready for Review", "#F59E0B"))
             review_fhir_values = gr.HTML("<span style='font-family:JetBrains Mono,monospace;'>FHIR values appear here.</span>")
-            section_one_text = gr.Textbox(label="NHS Clinic Letter", lines=20, interactive=True)
+            section_one_text = gr.Textbox(label="Document", lines=20, interactive=True)
             section_two_text = gr.Textbox(label="Section 2", lines=5, interactive=True, visible=False)
             section_three_text = gr.Textbox(label="Section 3", lines=5, interactive=True, visible=False)
             section_four_text = gr.Textbox(label="Section 4", lines=5, interactive=True, visible=False)
@@ -908,20 +1107,40 @@ def build_ui() -> gr.Blocks:
             download_text_file = gr.File(label="Download as Text", visible=False)
             hidden_copy_button = gr.Button("hidden-copy", visible=True, elem_id="hidden-copy")
             hidden_download_button = gr.Button("hidden-download", visible=True, elem_id="hidden-download")
-            gr.HTML("""<div style='display:flex;gap:12px;margin-top:24px;justify-content:center;'><button onclick=\"(function(){var el=document.getElementById('signed-letter-text');var text='';if(el){text=el.innerText||el.textContent;}if(!text){document.querySelectorAll('textarea').forEach(function(t){if(t.value&&t.value.length>50)text=t.value;});}if(!text){alert('No letter text found');return;}try{navigator.clipboard.writeText(text.trim()).then(function(){alert('Copied to clipboard!');});}catch(e){var ta=document.createElement('textarea');ta.value=text.trim();document.body.appendChild(ta);ta.select();document.execCommand('copy');document.body.removeChild(ta);alert('Copied to clipboard!');}})()\" style='background:transparent; color:#1A1A2E; border:2px solid #D4AF37; padding:12px 24px; border-radius:8px; font-family:'Inter',sans-serif; font-weight:600; font-size:14px; cursor:pointer; transition:all 0.3s ease;' onmouseover=\"this.style.background='rgba(212,175,55,0.1)';this.style.boxShadow='0 0 12px rgba(212,175,55,0.3)';this.style.transform='translateY(-2px)'\" onmouseout=\"this.style.background='transparent';this.style.boxShadow='none';this.style.transform='translateY(0)'\">📋 Copy to Clipboard</button><button onclick=\"(function(){console.log('Clarke: Download clicked');var el=document.getElementById('signed-letter-text');var text='';if(el){text=el.innerText||el.textContent;}if(!text){document.querySelectorAll('textarea').forEach(function(t){if(t.value&&t.value.length>50)text=t.value;});}if(!text){alert('No letter text found');return;}var a=document.createElement('a');a.href='data:text/plain;charset=utf-8,'+encodeURIComponent(text.trim());a.download='clinic_letter.txt';a.style.display='none';document.body.appendChild(a);a.click();document.body.removeChild(a);console.log('Clarke: Download complete via data URI');})()\" style='background:transparent; color:#1A1A2E; border:2px solid #D4AF37; padding:12px 24px; border-radius:8px; font-family:'Inter',sans-serif; font-weight:600; font-size:14px; cursor:pointer; transition:all 0.3s ease;' onmouseover=\"this.style.background='rgba(212,175,55,0.1)';this.style.boxShadow='0 0 12px rgba(212,175,55,0.3)';this.style.transform='translateY(-2px)'\" onmouseout=\"this.style.background='transparent';this.style.boxShadow='none';this.style.transform='translateY(0)'\">📄 Download as Text</button></div>""")
+            gr.HTML("""<div style='display:flex;gap:12px;margin-top:24px;justify-content:center;'><button onclick=\"(function(){var el=document.getElementById('signed-letter-text');var text='';if(el){text=el.innerText||el.textContent;}if(!text){document.querySelectorAll('textarea').forEach(function(t){if(t.value&&t.value.length>50)text=t.value;});}if(!text){alert('No letter text found');return;}try{navigator.clipboard.writeText(text.trim()).then(function(){alert('Copied to clipboard!');});}catch(e){var ta=document.createElement('textarea');ta.value=text.trim();document.body.appendChild(ta);ta.select();document.execCommand('copy');document.body.removeChild(ta);alert('Copied to clipboard!');}})()\" style='background:transparent; color:#1A1A2E; border:2px solid #D4AF37; padding:12px 24px; border-radius:8px; font-family:Inter,sans-serif; font-weight:600; font-size:14px; cursor:pointer; transition:all 0.3s ease;' onmouseover=\"this.style.background='rgba(212,175,55,0.1)';this.style.boxShadow='0 0 12px rgba(212,175,55,0.3)';this.style.transform='translateY(-2px)'\" onmouseout=\"this.style.background='transparent';this.style.boxShadow='none';this.style.transform='translateY(0)'\">📋 Copy to Clipboard</button><button onclick="clarkePrintPDF()" style='background:transparent; color:#1A1A2E; border:2px solid #D4AF37; padding:12px 24px; border-radius:8px; font-family:Inter,sans-serif; font-weight:600; font-size:14px; cursor:pointer; transition:all 0.3s ease;' onmouseover="this.style.background='rgba(212,175,55,0.1)';this.style.boxShadow='0 0 12px rgba(212,175,55,0.3)';this.style.transform='translateY(-2px)'" onmouseout="this.style.background='transparent';this.style.boxShadow='none';this.style.transform='translateY(0)'">📑 Download as PDF</button><button onclick=\"(function(){console.log('Clarke: Download clicked');var el=document.getElementById('signed-letter-text');var text='';if(el){text=el.innerText||el.textContent;}if(!text){document.querySelectorAll('textarea').forEach(function(t){if(t.value&&t.value.length>50)text=t.value;});}if(!text){alert('No letter text found');return;}var a=document.createElement('a');a.href='data:text/plain;charset=utf-8,'+encodeURIComponent(text.trim());a.download='clinic_letter.txt';a.style.display='none';document.body.appendChild(a);a.click();document.body.removeChild(a);console.log('Clarke: Download complete via data URI');})()\" style='background:transparent; color:#1A1A2E; border:2px solid #D4AF37; padding:12px 24px; border-radius:8px; font-family:Inter,sans-serif; font-weight:600; font-size:14px; cursor:pointer; transition:all 0.3s ease;' onmouseover=\"this.style.background='rgba(212,175,55,0.1)';this.style.boxShadow='0 0 12px rgba(212,175,55,0.3)';this.style.transform='translateY(-2px)'\" onmouseout=\"this.style.background='transparent';this.style.boxShadow='none';this.style.transform='translateY(0)'\">📄 Download as Text</button></div>""")
             gr.HTML("""<div style='position:sticky; bottom:0; left:0; right:0; z-index:100;'><button onclick=\"(function(){var el=document.getElementById('hidden-next-patient');if(!el){console.error('Clarke: hidden-next-patient not found');return;}if(el.tagName==='BUTTON'){el.click();}else{var b=el.querySelector('button');if(b)b.click();}console.log('Clarke: Next Patient clicked');})()\" style='display:block; width:100%; padding:18px 0; border:none; cursor:pointer; background:linear-gradient(135deg, #D4AF37 0%, #F0D060 100%); color:#1A1A2E; font-family:'Inter',sans-serif; font-weight:700; font-size:16px; letter-spacing:0.5px; transition:all 0.3s ease; box-shadow:0 -4px 16px rgba(212,175,55,0.3);' onmouseover=\"this.style.background='linear-gradient(135deg,#E8C84A,#F5E070)';this.style.boxShadow='0 -4px 24px rgba(212,175,55,0.5)';this.style.transform='translateY(-1px)'\" onmouseout=\"this.style.background='linear-gradient(135deg,#D4AF37,#F0D060)';this.style.boxShadow='0 -4px 16px rgba(212,175,55,0.3)';this.style.transform='translateY(0)'\">Next Patient →</button></div>""")
             hidden_next_patient_btn = gr.Button("hidden-next-patient", visible=True, elem_id="hidden-next-patient")
 
         with gr.Column(visible=True) as screen_s1:
             dashboard_html = gr.HTML(build_dashboard_html(clinic_payload))
+            doc_type_radio = gr.Radio(
+                choices=["Clinic Letter", "Ward Round Note"],
+                value="Clinic Letter",
+                label="📋  Document Type",
+                interactive=True,
+                elem_id="clarke-doc-type",
+            )
+            with gr.Accordion("⚙  Letter Preferences", open=False, elem_id="clarke-letter-prefs"):
+                gr.HTML("<p style='font-family:Inter,sans-serif;font-size:13px;color:#888;margin:0 0 12px 0;font-style:italic;'>Customise the generated clinic letter template. Changes persist for all patients in this clinic list.</p>")
+                with gr.Row():
+                    pref_clinician_name = gr.Textbox(label="Clinician Name", value="Dr Sarah Chen", interactive=True, scale=1)
+                    pref_clinician_title = gr.Textbox(label="Title / Role", value="Consultant, General Practice", interactive=True, scale=1)
+                with gr.Row():
+                    pref_hospital = gr.Textbox(label="Hospital / Trust", value="Clarke NHS Trust", interactive=True, scale=1)
+                    pref_department = gr.Textbox(label="Department", value="General Practice Department", interactive=True, scale=1)
+                with gr.Row():
+                    pref_gp_name = gr.Textbox(label="Addressee Name", value="Dr Andrew Wilson", interactive=True, scale=1)
+                    pref_signoff = gr.Textbox(label="Sign-off Phrase", value="Warm regards", interactive=True, scale=1)
+                pref_gp_address = gr.Textbox(label="Addressee Address", value="Riverside Medical Practice\n14 Harcourt Street\nLondon", lines=3, interactive=True)
+            gr.HTML(_letter_prefs_persistence_js())
             hidden_patient_buttons: list[gr.Button] = []
             for i in range(5):
                 hidden_patient_buttons.append(gr.Button(f"hidden-select-{i}", elem_id=f"hidden-select-{i}", visible=True))
 
         for i, hidden_btn in enumerate(hidden_patient_buttons):
             hidden_btn.click(
-                fn=lambda state, idx=i: _handle_patient_selection(state, idx),
-                inputs=[app_state],
+                fn=lambda state, cn, ct, ho, de, gp, so, ga, dt, idx=i: _handle_patient_selection(state, idx, cn, ct, ho, de, gp, so, ga, dt),
+                inputs=[app_state, pref_clinician_name, pref_clinician_title, pref_hospital, pref_department, pref_gp_name, pref_signoff, pref_gp_address, doc_type_radio],
                 outputs=[app_state, feedback_text, context_screen_html, section_one_text, section_two_text, section_three_text, section_four_text, signed_letter_html, review_fhir_values, screen_s1, screen_s2, screen_s3, screen_s4, screen_s5, screen_s6],
                 show_progress="full",
             )
@@ -936,6 +1155,6 @@ def build_ui() -> gr.Blocks:
         hidden_sign_off_btn.click(_sign_off_document, inputs=[app_state, section_one_text, section_two_text, section_three_text, section_four_text], outputs=[app_state, feedback_text, signed_letter_html, copy_to_clipboard_text, download_text_file, screen_s1, screen_s2, screen_s3, screen_s4, screen_s5, screen_s6], show_progress="full")
         hidden_copy_button.click(_copy_signed_document, inputs=[app_state], outputs=[app_state, feedback_text, copy_to_clipboard_text], show_progress="hidden")
         hidden_download_button.click(_prepare_signed_download, inputs=[app_state], outputs=[app_state, feedback_text, download_text_file], show_progress="hidden")
-        hidden_next_patient_btn.click(_next_patient, inputs=[app_state], outputs=[app_state, feedback_text, section_one_text, section_two_text, section_three_text, section_four_text, signed_letter_html, copy_to_clipboard_text, dashboard_html, screen_s1, screen_s2, screen_s3, screen_s4, screen_s5, screen_s6], show_progress="hidden")
+        hidden_next_patient_btn.click(_next_patient, inputs=[app_state, pref_clinician_name, pref_clinician_title, pref_hospital, pref_department, pref_gp_name, pref_signoff, pref_gp_address, doc_type_radio], outputs=[app_state, feedback_text, section_one_text, section_two_text, section_three_text, section_four_text, signed_letter_html, copy_to_clipboard_text, dashboard_html, screen_s1, screen_s2, screen_s3, screen_s4, screen_s5, screen_s6, pref_clinician_name, pref_clinician_title, pref_hospital, pref_department, pref_gp_name, pref_signoff, pref_gp_address], show_progress="hidden")
 
     return demo