Claude 3.5 Sonnet Hybrid RAG with ColBERT & Cross-Encoder Reranking

You are a Principal AI Information Retrieval Specialist and Staff ML Systems Engineer.

Architect and implement a high-precision, two-stage Hybrid Retrieval-Augmented Generation (RAG) system in {Python 3.12 with LangChain / LlamaIndex & HuggingFace Transformers} for {Enterprise Financial Regulatory Filing (SEC 10-K, 10-Q) Analysis & Compliance Audit}.

### RETRIEVAL ARCHITECTURAL MANDATES:
1. **Stage 1: Multi-Modal First-Pass Retrieval (High Recall)**
   - Combine lexical BM25 keyword search with dense vector semantic search (OpenAI text-embedding-3-large or Cohere embed-v3).
   - Merge top-100 candidate pools using Reciprocal Rank Fusion (RRF) with configurable k-constant.
2. **Stage 2: Second-Pass Deep Reranking (High Precision)**
   - Apply {Cross-Encoder (BAAI/bge-reranker-large on GPU)} (e.g. BAAI/bge-reranker-large, Cohere Rerank v3, or ColBERTv2 late-interaction token alignment).
   - Score exact query-document token cross-attention, filtering out irrelevant semantic distractors.
3. **Stage 3: Dynamic Context Compression & Window Packing**
   - Pack only top-K reranked chunks into Claude 3.5 Sonnet's context window, preserving source metadata citations (`[Doc #, Page #]`).
4. **Hallucination Guardrails**: Enforce strict grounding prompts requiring Claude to cite specific retrieved sentence spans.

Vector Database & Indexing:
{Qdrant with Dense Vectors + BM25 Sparse Inverted Index}

Document Corpus & Chunking:
{Semantic section boundary chunking (512 tokens with 64 token overlap) with table Markdown preservation}