Update src/streamlit_app.py

src/streamlit_app.py

@@ -23,9 +23,6 @@ MAX_NEW_TOKENS = 192
 TOP_K_SEARCH = 3
 LOGO_URL = "https://teapotai.com/assets/logo.gif"
 
-# How many (user,assistant) pairs to include in the prompt by default
-MAX_TURNS_IN_PROMPT = 6
-
 st.set_page_config(page_title="TeapotAI Chat", page_icon="🫖", layout="centered")
 
 
@@ -68,14 +65,17 @@ if "needs_answer" not in st.session_state:
 
 
 # =========================
-# HEADER (prevent logo flash)
+# HEADER (reduce logo flash)
+# - fixed width prevents "giant image then shrink"
+# - container keeps layout stable
 # =========================
-col1, col2 = st.columns([1, 7], vertical_alignment="center")
-with col1:
-    st.image(LOGO_URL, width=56)
-with col2:
-    st.markdown("## TeapotAI Chat")
-    st.caption("Grounded answers with web context")
+with st.container():
+    col1, col2 = st.columns([1, 7], vertical_alignment="center")
+    with col1:
+        st.image(LOGO_URL, width=56)  # fixed width = stable, no huge flash
+    with col2:
+        st.markdown("## TeapotAI Chat")
+        st.caption("Grounded answers with web context")
 
 
 # =========================
@@ -102,14 +102,6 @@ with st.sidebar:
         placeholder="Extra context appended after web snippets…",
     )
 
-    max_turns = st.slider(
-        "Conversation turns in prompt",
-        min_value=0,
-        max_value=12,
-        value=MAX_TURNS_IN_PROMPT,
-        help="How many recent (user, assistant) pairs to include in the prompt.",
-    )
-
 
 # =========================
 # WEB SEARCH (ALWAYS ON)
@@ -145,64 +137,28 @@ def web_search_snippets(query: str):
 
 
 # =========================
-# CONTEXT + PROMPT BUILDING
+# PROMPT BUILDING (NO CONVERSATION HISTORY)
 # =========================
 def count_tokens(text: str) -> int:
     return len(tokenizer.encode(text)) if text else 0
 
 
-def build_conversation(messages, turns: int) -> str:
-    """
-    Build a compact transcript from the last `turns` (user,assistant) pairs.
-    """
-    if turns <= 0:
-        return ""
-
-    # Collect last 2*turns messages ending at the most recent message
-    # Keep only user/assistant roles.
-    filtered = [m for m in messages if m.get("role") in ("user", "assistant")]
-
-    # Take tail, but ensure we start on a user message if possible
-    tail = filtered[-(2 * turns) :]
-    # If first is assistant, drop it (misaligned pair)
-    if tail and tail[0]["role"] == "assistant":
-        tail = tail[1:]
-
-    lines = []
-    for m in tail:
-        role = "User" if m["role"] == "user" else "Assistant"
-        content = (m.get("content") or "").strip()
-        if content:
-            lines.append(f"{role}: {content}")
-    return "\n".join(lines).strip()
-
-
-def truncate_to_token_budget(full_prompt: str, max_tokens: int) -> str:
-    ids = tokenizer.encode(full_prompt)
+def truncate_to_token_budget(text: str, max_tokens: int) -> str:
+    ids = tokenizer.encode(text)
     if len(ids) <= max_tokens:
-        return full_prompt
-    # Tail truncate to keep the most recent instruction + question
-    ids = ids[-max_tokens:]
+        return text
+    ids = ids[-max_tokens:]  # tail truncate
     return tokenizer.decode(ids, skip_special_tokens=True)
 
 
-def build_prompt(web_ctx: str, local_ctx: str, system: str, convo: str, question: str) -> str:
-    # Order matters: context first, then system, then convo, then question.
-    parts = []
+def build_prompt(web_ctx: str, local_ctx: str, system: str, question: str) -> str:
     ctx = f"{web_ctx}\n\n{local_ctx}".strip()
+    parts = []
     if ctx:
-        parts.append("Context:\n" + ctx)
-
-    parts.append("System:\n" + system.strip())
-
-    if convo:
-        parts.append("Conversation:\n" + convo)
-
-    parts.append("User:\n" + question.strip())
-    parts.append("Assistant:\n")  # encourages continuation style
-
+        parts.append(ctx)
+    parts.append(system.strip())
+    parts.append(question.strip())
     raw = "\n\n".join(parts).strip() + "\n"
-    # Enforce input budget at token level
     return truncate_to_token_budget(raw, MAX_INPUT_TOKENS)
 
 
@@ -237,9 +193,38 @@ def handle_feedback(idx: int):
     val = st.session_state.get(f"fb_{idx}")
     st.session_state.messages[idx]["feedback"] = val
 
-    # If you later add LangSmith run_ids per message, hook it here.
-    # (Keeping this simple/stable like your previous version.)
-    # if ls_client and st.session_state.messages[idx].get("run_id"): ...
+    # If you want to wire LangSmith feedback to a run later, store run_id per message and use it here.
+    # For now we keep it stable and local like the earlier version.
+
+
+def render_inline_controls(msg: dict, feedback_key: str, feedback_disabled: bool, feedback_idx: int | None):
+    """
+    Inline row under assistant message:
+      ℹ️ popover (full prompt), thumbs, metrics
+    """
+    c_info, c_fb, c_metrics = st.columns([1.1, 1.7, 10], vertical_alignment="center")
+
+    with c_info:
+        with st.popover("ℹ️", help="Inspect"):
+            st.markdown("**Prompt (sent to model)**")
+            st.code(msg.get("prompt", ""), language="text")
+
+    with c_fb:
+        st.feedback(
+            "thumbs",
+            key=feedback_key,
+            disabled=feedback_disabled,
+            on_change=(handle_feedback if (not feedback_disabled and feedback_idx is not None) else None),
+            args=((feedback_idx,) if (not feedback_disabled and feedback_idx is not None) else None),
+        )
+
+    with c_metrics:
+        st.caption(
+            f"🔎 {msg.get('search_time', 0.0):.2f}s "
+            f"• 🧠 {msg.get('gen_time', 0.0):.2f}s "
+            f"• ⚡ {msg.get('tps', 0.0):.1f} tok/s "
+            f"• 🧾 in {msg.get('input_tokens', 0)} • out {msg.get('output_tokens', 0)}"
+        )
 
 
 # =========================
@@ -250,32 +235,16 @@ for i, msg in enumerate(st.session_state.messages):
         st.markdown(msg["content"])
 
         if msg["role"] == "assistant":
-            # Inline row: info popover + thumbs + metrics
-            c_info, c_fb, c_metrics = st.columns([1.2, 1.4, 10], vertical_alignment="center")
-
-            with c_info:
-                with st.popover("ℹ️", help="Inspect prompt"):
-                    st.markdown("**Prompt sent to model**")
-                    st.code(msg.get("prompt", ""), language="text")
-
-            with c_fb:
-                key = f"fb_{i}"
-                st.session_state.setdefault(key, msg.get("feedback"))
-                st.feedback(
-                    "thumbs",
-                    key=key,
-                    disabled=msg.get("feedback") is not None,
-                    on_change=handle_feedback,
-                    args=(i,),
-                )
-
-            with c_metrics:
-                st.caption(
-                    f"🔎 {msg['search_time']:.2f}s "
-                    f"• 🧠 {msg['gen_time']:.2f}s "
-                    f"• ⚡ {msg['tps']:.1f} tok/s "
-                    f"• 🧾 in {msg['input_tokens']} • out {msg['output_tokens']}"
-                )
+            # Ensure feedback state exists
+            k = f"fb_{i}"
+            st.session_state.setdefault(k, msg.get("feedback"))
+
+            render_inline_controls(
+                msg=msg,
+                feedback_key=k,
+                feedback_disabled=(msg.get("feedback") is not None),
+                feedback_idx=i,
+            )
 
 
 # =========================
@@ -299,25 +268,16 @@ if (
 ):
     question = st.session_state.messages[-1]["content"]
 
-    # Web search
     web_ctx, search_time = web_search_snippets(question)
 
-    # Conversation transcript (from prior messages, excluding current user msg is fine either way;
-    # keeping it includes the last user msg too, but we also add question explicitly.)
-    convo = build_conversation(st.session_state.messages[:-1], turns=max_turns)
-
-    # Prompt
     prompt = build_prompt(
         web_ctx=web_ctx,
         local_ctx=local_context,
         system=system_prompt,
-        convo=convo,
         question=question,
     )
-
     input_tokens = count_tokens(prompt)
 
-    # Stream normally
     with st.chat_message("assistant"):
         placeholder = st.empty()
         start = time.perf_counter()
@@ -331,27 +291,24 @@ if (
         output_tokens = count_tokens(final_text)
         tps = output_tokens / gen_time if gen_time > 0 else 0.0
 
-        # Inline row under the live message
-        c_info, c_fb, c_metrics = st.columns([1.2, 1.4, 10], vertical_alignment="center")
-
-        with c_info:
-            with st.popover("ℹ️", help="Inspect prompt"):
-                st.markdown("**Prompt sent to model**")
-                st.code(prompt, language="text")
-
-        # For the live message, we don't have a saved index yet; show disabled thumbs placeholder
-        with c_fb:
-            st.feedback("thumbs", key="fb_live", disabled=True)
+        live_msg = {
+            "prompt": prompt,
+            "search_time": search_time,
+            "gen_time": gen_time,
+            "input_tokens": input_tokens,
+            "output_tokens": output_tokens,
+            "tps": tps,
+        }
 
-        with c_metrics:
-            st.caption(
-                f"🔎 {search_time:.2f}s "
-                f"• 🧠 {gen_time:.2f}s "
-                f"• ⚡ {tps:.1f} tok/s "
-                f"• 🧾 in {input_tokens} • out {output_tokens}"
-            )
+        # Inline controls for the live message (thumbs disabled until it’s saved)
+        render_inline_controls(
+            msg=live_msg,
+            feedback_key="fb_live",
+            feedback_disabled=True,
+            feedback_idx=None,
+        )
 
-    # Persist assistant message (so feedback attaches properly after rerun)
+    # Save assistant message so thumbs attach after rerun
     st.session_state.messages.append(
         {
             "role": "assistant",