abdessamadr/bank-rag-mcp

How abdessamadr/bank-rag-mcp works

Bank RAG MCP extracts bank annual reports, indexes page-aware chunks, retrieves relevant passages, and generates French answers with source citations. The main workflow is PDF extraction → chunking → embedding and storage → hybrid retrieval → answer generation. Evaluation and vector-store auditing are additional capabilities. The README documents an MCP server, but no corresponding source directory appears in the supplied tree, so that integration is shown as unverified; the excerpts do not establish a runtime connection to it.

Main entry points: retriever.py Hybrid retriever · answer_generator.py Answer generator · data Chroma vector store · embedder.py Embedding indexer

Components 18

Document ingestion

PDF extraction extract_pdfs.py
Extracted pages
Page chunking src/chunking/chunker.py
Page-aware chunks

Search and storage

Embedding indexer src/ingestion/embedder.py
E5 embeddings
Chroma vector store data
Hybrid retriever src/retrieval/retriever.py

Answer generation

Answer generator src/generation/answer_generator.py
Cited answer src/generation/answer_generator.py
MCP search server

Evaluation and audit

Retriever evaluation src/retrieval/evaluate_retriever.py
Embedding benchmark src/evaluation/benchmark_embeddings.py
Model configurations src/evaluation/model_config.py
Vector-store audit src/audit/audit_vectorstore.py

Other components

Questioner
Bank report PDFs
Ollama LLM

Connections 18

Questioner → Answer generator: asks question src/generation/answer_generator.py
Bank report PDFs → PDF extraction: provides PDFs extract_pdfs.py
PDF extraction → Extracted pages: writes pages extract_pdfs.py
Extracted pages → Page chunking: reads pages src/chunking/chunker.py
Page chunking → Page-aware chunks: writes chunks src/chunking/chunker.py
Page-aware chunks → Embedding indexer: loads chunks src/ingestion/embedder.py
Embedding indexer → E5 embeddings: encodes passages src/ingestion/embedder.py
Embedding indexer → Chroma vector store: indexes chunks src/ingestion/embedder.py
Hybrid retriever → E5 embeddings: encodes queries src/retrieval/retriever.py
Hybrid retriever → Chroma vector store: queries and reads src/retrieval/retriever.py
Answer generator → Hybrid retriever: searches passages src/generation/answer_generator.py
Answer generator → Ollama LLM: sends context src/generation/answer_generator.py
Answer generator → Cited answer: returns citations src/generation/answer_generator.py
Retriever evaluation → Hybrid retriever: evaluates vectors src/retrieval/evaluate_retriever.py
Embedding benchmark → Model configurations: loads settings src/evaluation/benchmark_embeddings.py
Embedding benchmark → Chroma vector store: builds and queries src/evaluation/benchmark_embeddings.py
Vector-store audit → Chroma vector store: checks collection src/audit/audit_vectorstore.py
MCP search server → Hybrid retriever: may expose search

Drawn by GitDiagram on November 11, 2026 from the repository's file tree, README and source files. Read as Markdown ·