Core

"""
Core: RAG Context Optimizer & Precision Re-Ranker
Prunes noisy embedding vectors and packs maximum semantic context into prompts.
"""
from typing import List, Dict, Any
from pydantic import BaseModel, Field

class RAGChunk(BaseModel):
    chunk_id: str
    content: str
    relevance_score: float
    token_count: int

class OptimizedContextEnvelope(BaseModel):
    pruned_chunks_count: int
    injected_token_count: int
    compression_ratio: float
    context_text: str

class CoreRAGOptimizer:
    def __init__(self, max_context_tokens: int = 2048, score_threshold: float = 0.72):
        self.max_context_tokens = max_context_tokens
        self.score_threshold = score_threshold

    def optimize_retrieval(self, raw_chunks: List[RAGChunk]) -> OptimizedContextEnvelope:
        # Sort by relevance score descending (cross-encoder rank)
        filtered = [c for c in raw_chunks if c.relevance_score >= self.score_threshold]
        filtered.sort(key=lambda x: x.relevance_score, reverse=True)

        selected_chunks: List[str] = []
        current_tokens = 0

        for chunk in filtered:
            if current_tokens + chunk.token_count <= self.max_context_tokens:
                selected_chunks.append(chunk.content)
                current_tokens += chunk.token_count
            else:
                break

        original_tokens = sum(c.token_count for c in raw_chunks)
        compression = 1.0 - (current_tokens / max(1, original_tokens))

        return OptimizedContextEnvelope(
            pruned_chunks_count=len(raw_chunks) - len(selected_chunks),
            injected_token_count=current_tokens,
            compression_ratio=round(compression, 2),
            context_text="\n---\n".join(selected_chunks)
        )