inference.py

"""JEV-2 Inference Engine with Typed Decision Primitives (Noul, Choice, Score)
Compatible with VTXAI/VTX-JEV-2 and VTXAI/VTX-MOD-1.

Usage:
    from inference import JevClient, Choice, Noul, Score

    client = JevClient.from_pretrained("VTXAI/VTX-JEV-2")

    response = client.system_one(
        state="I was charged twice and production is unavailable.",
        questions={
            "refund": Noul("Does the customer request a refund?"),
            "team": Choice(
                "Which team should handle this?",
                {"billing": "Payments", "technical": "Production outage"},
            ),
            "severity": Score(
                "How severe is the impact?",
                ["Minor", "Major", "Critical"],
            ),
        },
    )

    print(response.nouls["refund"].noul)       # bool (e.g. True)
    print(response.choices["team"].choice)     # key (e.g. 'billing')
    print(response.scores["severity"].score)   # top label or expected rank
    print(response.to_dict())
"""

from __future__ import annotations

import os
import json
import time
from dataclasses import dataclass, field
from pathlib import Path
from typing import Dict, List, Union, Any, Optional

import numpy as np
from safetensors.numpy import load_file
from tokenizers import Tokenizer
from huggingface_hub import hf_hub_download

# --- Decision Primitives ---

class Noul:
    """Binary decision / Gating condition (Yes / No)."""
    def __init__(self, question: str, positive_anchor: str = "yes positive true confirm affirmative request", negative_anchor: str = "no negative false ignore denial safe", threshold: float = 0.5):
        self.question = question
        self.positive_anchor = positive_anchor
        self.negative_anchor = negative_anchor
        self.threshold = threshold

class Choice:
    """Categorical routing / Selection among options."""
    def __init__(self, question: str, options: Union[List[str], Dict[str, str]]):
        self.question = question
        if isinstance(options, list):
            self.options = {opt: opt for opt in options}
        else:
            self.options = options

class Score:
    """Ordinal priority / Severity calibration."""
    def __init__(self, question: str, scale: List[str]):
        self.question = question
        self.scale = scale

# --- Response Containers ---

@dataclass
class NoulResult:
    noul: bool
    probability: float
    confidence: float

    def to_dict(self) -> Dict[str, Any]:
        return {
            "noul": self.noul,
            "probability": round(self.probability, 4),
            "confidence": round(self.confidence, 4),
        }

@dataclass
class ChoiceResult:
    choice: str
    confidence: float
    distribution: Dict[str, float]

    def to_dict(self) -> Dict[str, Any]:
        return {
            "choice": self.choice,
            "confidence": round(self.confidence, 4),
            "distribution": {k: round(v, 4) for k, v in self.distribution.items()},
        }

@dataclass
class ScoreResult:
    score: str
    expected_index: float
    distribution: Dict[str, float]

    def to_dict(self) -> Dict[str, Any]:
        return {
            "score": self.score,
            "expected_index": round(self.expected_index, 4),
            "distribution": {k: round(v, 4) for k, v in self.distribution.items()},
        }

@dataclass
class SystemOneResponse:
    nouls: Dict[str, NoulResult] = field(default_factory=dict)
    choices: Dict[str, ChoiceResult] = field(default_factory=dict)
    scores: Dict[str, ScoreResult] = field(default_factory=dict)
    latency_ms: float = 0.0

    def to_dict(self) -> Dict[str, Any]:
        return {
            "nouls": {k: v.to_dict() for k, v in self.nouls.items()},
            "choices": {k: v.to_dict() for k, v in self.choices.items()},
            "scores": {k: v.to_dict() for k, v in self.scores.items()},
            "latency_ms": round(self.latency_ms, 3),
        }

# --- JevClient Core ---

class JevClient:
    """JEV-2 Ultra-fast System 1 Decision & Routing Client."""

    def __init__(self, weights: np.ndarray, tokenizer: Tokenizer, config: dict):
        self.weights = weights  # (V, D) normalized token embeddings
        self.tokenizer = tokenizer
        self.config = config
        self.dim = weights.shape[1]
        self.vocab_size = weights.shape[0]

    @classmethod
    def from_pretrained(cls, model_name_or_path: str) -> "JevClient":
        """Load JevClient from local folder or Hugging Face Hub."""
        path = Path(model_name_or_path)
        if path.exists() and (path / "model.safetensors").exists():
            model_file = str(path / "model.safetensors")
            tok_file = str(path / "tokenizer.json")
            cfg_file = str(path / "config.json") if (path / "config.json").exists() else None
        else:
            model_file = hf_hub_download(model_name_or_path, "model.safetensors")
            tok_file = hf_hub_download(model_name_or_path, "tokenizer.json")
            try:
                cfg_file = hf_hub_download(model_name_or_path, "config.json")
            except Exception:
                cfg_file = None

        tensors = load_file(model_file)
        weight_key = "embeddings" if "embeddings" in tensors else list(tensors.keys())[0]
        weights = tensors[weight_key].astype(np.float32)

        # Normalize token embedding table once
        norms = np.linalg.norm(weights, axis=-1, keepdims=True)
        weights = weights / np.maximum(norms, 1e-12)

        tokenizer = Tokenizer.from_file(tok_file)
        cfg = json.loads(Path(cfg_file).read_text()) if cfg_file else {}

        return cls(weights, tokenizer, cfg)

    def encode(self, texts: List[str]) -> np.ndarray:
        """Fast vectorized token lookup and mean pooling on CPU."""
        encoded = self.tokenizer.encode_batch(texts)
        res = np.zeros((len(texts), self.dim), dtype=np.float32)

        for i, item in enumerate(encoded):
            ids = [tid for tid in item.ids if 0 <= tid < self.vocab_size]
            if ids:
                tok_vecs = self.weights[ids]
                # Mean pool
                pooled = np.mean(tok_vecs, axis=0)
                norm = np.linalg.norm(pooled)
                if norm > 1e-9:
                    res[i] = pooled / norm
            else:
                res[i] = 0.0
        return res

    def system_one(self, state: str, questions: Dict[str, Union[Noul, Choice, Score]]) -> SystemOneResponse:
        """Execute typed System 1 non-autoregressive decision primitives concurrently."""
        t0 = time.perf_counter()

        response = SystemOneResponse()

        # Gather texts to encode in a single optimized pass
        batch_texts = [state]
        query_map = {}

        for key, q in questions.items():
            if isinstance(q, Noul):
                # Combined query, positive anchor, negative anchor
                idx_q = len(batch_texts)
                batch_texts.append(f"{q.question} {state}")
                idx_pos = len(batch_texts)
                batch_texts.append(f"{q.question} {q.positive_anchor}")
                idx_neg = len(batch_texts)
                batch_texts.append(f"{q.question} {q.negative_anchor}")
                query_map[key] = ("noul", q, (idx_q, idx_pos, idx_neg))

            elif isinstance(q, Choice):
                idx_start = len(batch_texts)
                keys = list(q.options.keys())
                for k in keys:
                    batch_texts.append(f"{q.question} {q.options[k]}")
                query_map[key] = ("choice", q, (idx_start, keys))

            elif isinstance(q, Score):
                idx_start = len(batch_texts)
                for opt in q.scale:
                    batch_texts.append(f"{q.question} {opt}")
                query_map[key] = ("score", q, (idx_start, q.scale))

        # Single batch CPU encode
        all_vecs = self.encode(batch_texts)
        state_vec = all_vecs[0]

        # Process each primitive result
        for key, info in query_map.items():
            q_type = info[0]

            if q_type == "noul":
                _, q, (idx_q, idx_pos, idx_neg) = info
                q_vec = all_vecs[idx_q]
                pos_vec = all_vecs[idx_pos]
                neg_vec = all_vecs[idx_neg]

                sim_pos = float(np.dot(q_vec, pos_vec))
                sim_neg = float(np.dot(q_vec, neg_vec))
                diff = sim_pos - sim_neg
                prob = float(1.0 / (1.0 + np.exp(-diff * 12.0)))
                decision = bool(prob >= q.threshold)
                conf = float(prob if decision else (1.0 - prob))
                response.nouls[key] = NoulResult(noul=decision, probability=prob, confidence=conf)

            elif q_type == "choice":
                _, q, (idx_start, keys) = info
                cand_vecs = all_vecs[idx_start : idx_start + len(keys)]
                # Affinities with state and question context
                sims = np.dot(cand_vecs, state_vec)
                # Softmax with temperature
                exp_s = np.exp(sims * 15.0)
                probs = exp_s / np.sum(exp_s)
                best_idx = int(np.argmax(probs))
                response.choices[key] = ChoiceResult(
                    choice=keys[best_idx],
                    confidence=float(probs[best_idx]),
                    distribution={k: float(p) for k, p in zip(keys, probs)},
                )

            elif q_type == "score":
                _, q, (idx_start, scale) = info
                cand_vecs = all_vecs[idx_start : idx_start + len(scale)]
                sims = np.dot(cand_vecs, state_vec)
                exp_s = np.exp(sims * 15.0)
                probs = exp_s / np.sum(exp_s)
                best_idx = int(np.argmax(probs))
                exp_idx = float(np.sum(np.arange(len(scale)) * probs))
                response.scores[key] = ScoreResult(
                    score=scale[best_idx],
                    expected_index=exp_idx,
                    distribution={k: float(p) for k, p in zip(scale, probs)},
                )

        response.latency_ms = (time.perf_counter() - t0) * 1000
        return response