How mayankpuvvala/financial_rag works
Financial RAG answers questions about SEC 10-K filings with grounded, numbered citations. A question enters through the API or Python/CLI interface, is classified and resolved (including on-demand filing ingestion), then follows direct retrieval or multi-document/temporal decomposition, retrieval, and answer generation or synthesis. Filing ingestion downloads, parses, chunks, embeds, and indexes documents. The README documents these major capabilities; several implementation details are unsampled, so uncertain connections are omitted.
Main entry points: retriever.py Hybrid retriever · auto_ingest.py On-demand indexing · synthesizer.py Multi-query synthesizer · query.py Query pipeline
Components 23
Interfaces
- HTTP API
api/app.py - Query pipeline
query.py
Query routing
- Company resolver
routing/resolver.py - Query classifier
routing/classifier.py - Query decomposer
routing/decomposer.py
Filing ingestion
- SEC ticker registry
ingestion/registry.py - On-demand indexing
ingestion/auto_ingest.py - Filing downloader
ingestion/downloader.py - Filing parser
ingestion/parser.py - Parent-child chunker
ingestion/chunker.py
Retrieval and indexing
- Embedding indexer
ingestion/embedder.py - Hybrid retriever
retrieval/retriever.py - Cross-encoder reranker
retrieval/reranker.py - Parent section store
retrieval/parent_store.py - Qdrant collections
retrieval/vector_store.py
Answer generation
- Cited answer generator
generation/generator.py - Multi-query synthesizer
generation/synthesizer.py
Other components
- Questioner
- Settings
config.py - Domain models
models.py - SEC EDGAR
- Groq LLM
- Cited answer
Connections 32
- Questioner → HTTP API: asks via HTTP
README.md - Questioner → Query pipeline: asks via CLI
README.md - HTTP API → Query pipeline: invokes
api/app.py - Query pipeline → Company resolver: classifies and resolves
query.py - Company resolver → Query classifier: classifies
routing/resolver.py - Company resolver → SEC ticker registry: resolves ticker
routing/resolver.py - Company resolver → On-demand indexing: ensures indexed
routing/resolver.py - SEC ticker registry → SEC EDGAR: looks up tickers
README.md - Filing downloader → SEC EDGAR: downloads filings
README.md - On-demand indexing → Filing downloader: fetches filing
ingestion/auto_ingest.py - On-demand indexing → Filing parser: parses filing
ingestion/auto_ingest.py - On-demand indexing → Parent-child chunker: chunks sections
ingestion/auto_ingest.py - On-demand indexing → Embedding indexer: indexes chunks
ingestion/auto_ingest.py - Filing downloader → Filing parser: feeds filing records
README.md - Filing parser → Parent-child chunker: produces sections
README.md - Parent-child chunker → Embedding indexer: produces chunks
README.md - Embedding indexer → Qdrant collections: upserts vectors
ingestion/embedder.py - Query pipeline → Hybrid retriever: retrieves directly
query.py - Query pipeline → Multi-query synthesizer: handles multi-year queries
query.py - Multi-query synthesizer → Query decomposer: decomposes queries
generation/synthesizer.py - Multi-query synthesizer → Hybrid retriever: retrieves per question
generation/synthesizer.py - Multi-query synthesizer → Cited answer generator: generates sub-answers
generation/synthesizer.py - Hybrid retriever → Embedding indexer: encodes queries
retrieval/retriever.py - Hybrid retriever → Qdrant collections: hybrid searches
retrieval/retriever.py - Hybrid retriever → Cross-encoder reranker: reranks matches
retrieval/retriever.py - Hybrid retriever → Parent section store: loads parent text
retrieval/retriever.py - Hybrid retriever → Cited answer generator: supplies context
README.md - Cited answer generator → Groq LLM: requests completion
generation/generator.py - Query classifier → Groq LLM: classifies question
README.md - Query decomposer → Groq LLM: decomposes question
README.md - Cited answer generator → Cited answer: returns citations
README.md - Multi-query synthesizer → Cited answer: returns synthesis
generation/synthesizer.py