Spaces:
Sleeping
Sleeping
| """Validation node for DocIntel.""" | |
| from app.agents.state import AgentState | |
| from app.config import settings | |
| def validate_response(state: AgentState) -> AgentState: | |
| """Validate response and calculate confidence""" | |
| print("🔍 Validating response...") | |
| response = state.get("response", "") | |
| documents = state.get("documents", []) | |
| current_confidence = state.get("confidence", 0.0) | |
| if not response: | |
| state["validated"] = False | |
| state["confidence"] = 0.0 | |
| print("❌ No response to validate") | |
| return state | |
| # Factor 1: Document relevance (40% weight) | |
| doc_scores = [d.get("score", 0) for d in documents[:3]] | |
| avg_doc_score = sum(doc_scores) / len(doc_scores) if doc_scores else 0 | |
| # Factor 2: Response length (20% weight) - longer responses indicate more content | |
| length_score = min(1.0, len(response) / 500) # Cap at 500 chars | |
| # Factor 3: Number of sources (20% weight) | |
| source_score = min(1.0, len(documents) / settings.TOP_K) if documents else 0 | |
| # Factor 4: Confidence from retrieval (20% weight) | |
| retrieval_score = min(1.0, current_confidence * 2) # Scale up | |
| # Calculate weighted confidence | |
| confidence = ( | |
| avg_doc_score * 0.40 + | |
| length_score * 0.20 + | |
| source_score * 0.20 + | |
| retrieval_score * 0.20 | |
| ) | |
| # Cap at 0.95 | |
| confidence = min(0.95, confidence) | |
| # If no documents, low confidence | |
| if not documents: | |
| confidence = 0.1 | |
| state["confidence"] = confidence | |
| state["validated"] = confidence > 0.3 # Threshold for validation | |
| print(f"✅ Confidence: {confidence:.3f} (doc: {avg_doc_score:.2f}, len: {length_score:.2f}, src: {source_score:.2f}, ret: {retrieval_score:.2f})") | |
| return state |