RAG with Reranking — Hybrid Search Pipeline

## Task
RAG pipeline with hybrid search and reranking for high-precision Q&A.

## Requirements
- Vector DB: pgvector, Pinecone, or Qdrant
- Embeddings: OpenAI text-embedding-3-small or Cohere embed-v3
- Reranker: Cohere rerank or cross-encoder model
- Language: Python or TypeScript

## Pipeline
```
Query → [Hybrid Search] → [Rerank] → [LLM Generate]

1. Hybrid Search (parallel):
   a. Vector search: embed query → top 20 by cosine similarity
   b. Keyword search: BM25/FTS on same corpus → top 20
   c. Merge results using Reciprocal Rank Fusion (RRF)

2. Rerank:
   - Take merged top 30 results
   - Rerank with cross-encoder (query, document) pairs
   - Keep top 5

3. Generate:
   - Inject top 5 chunks as context
   - System prompt: "Answer based only on provided context"
   - Include source citations
```

## Implementation Notes
1. Chunk documents at 512 tokens with 50-token overlap
2. Store metadata: source URL, title, chunk index
3. Cache embeddings — don't re-embed on every query
4. Include "I don't have enough information" when context is insufficient
5. Return confidence score based on reranker scores