How nickr145/crag-finqa works
Financial CRAG answers FinQA financial questions through a hybrid retrieval and corrective-generation workflow. The main entry points are the FastAPI query endpoint and the CLI benchmark; both initialize the corpus, retriever, and LangGraph workflow. The graph retrieves documents, scores relevance, then refines internal context or searches SEC EDGAR before Claude generates an answer. The Streamlit demo and numeric-tolerance benchmark are additional capabilities. Edges reflect the sampled source and README; unsampled UI wiring is not inferred.
Main entry points: graph.py CRAG Graph · app.py Query API · main.py CLI Benchmark · loader.py Corpus Loader
Components 22
User Interfaces
- Query API
app.py - Streamlit Demo
frontend.py - CLI Benchmark
main.py
Corpus and Retrieval
- Corpus Loader
src/data/loader.py - Financial Parser
src/data/parser.py - Hybrid Retriever
src/retrieval/hybrid.py - EDGAR Search
src/retrieval/sec_edgar.py
CRAG Workflow
- CRAG Graph
src/workflow/graph.py - Workflow State
src/workflow/state.py - Retrieve
src/workflow/graph.py - Evaluate Relevance
src/workflow/graph.py - Document Scorer
src/evaluation/scorer.py - Route Action
src/workflow/graph.py - SEC Fallback
src/workflow/graph.py - Context Refinement
src/workflow/graph.py - Answer Generation
src/workflow/graph.py
Evaluation and Settings
- Numeric Tolerance
src/evaluation/metrics.py - Runtime Settings
config/settings.py
Other components
- Questioner
- FinQA Dataset
- Claude API
- SEC EDGAR
Connections 32
- Questioner → Query API: submits query
README.md - Query API → Corpus Loader: loads corpus
app.py - Query API → Hybrid Retriever: initializes
app.py - Query API → CRAG Graph: builds and invokes
app.py - Streamlit Demo → Query API: uses service
README.md - CLI Benchmark → Corpus Loader: loads corpus
main.py - CLI Benchmark → Hybrid Retriever: initializes
main.py - CLI Benchmark → CRAG Graph: builds and invokes
main.py - CLI Benchmark → Numeric Tolerance: scores answer
main.py - Corpus Loader → FinQA Dataset: fetches records
README.md - Corpus Loader → Financial Parser: formats records
src/data/loader.py - CRAG Graph → Workflow State: uses schema
src/workflow/graph.py - CRAG Graph → Retrieve: registers node
src/workflow/graph.py - CRAG Graph → Evaluate Relevance: registers node
src/workflow/graph.py - CRAG Graph → Route Action: registers routing
src/workflow/graph.py - CRAG Graph → SEC Fallback: registers node
src/workflow/graph.py - CRAG Graph → Context Refinement: registers node
src/workflow/graph.py - CRAG Graph → Answer Generation: registers node
src/workflow/graph.py - Retrieve → Hybrid Retriever: searches documents
src/workflow/graph.py - Retrieve → Evaluate Relevance: passes documents
src/workflow/graph.py - Evaluate Relevance → Document Scorer: scores documents
src/workflow/graph.py - Evaluate Relevance → Route Action: sets action
src/workflow/graph.py - Route Action → SEC Fallback: routes incorrect
src/workflow/graph.py - Route Action → Context Refinement: routes otherwise
src/workflow/graph.py - SEC Fallback → EDGAR Search: searches filings
src/workflow/graph.py - EDGAR Search → SEC EDGAR: queries and fetches
src/retrieval/sec_edgar.py - SEC Fallback → Answer Generation: passes context
src/workflow/graph.py - Context Refinement → Answer Generation: passes context
src/workflow/graph.py - Answer Generation → Claude API: requests answer
src/workflow/graph.py - Runtime Settings → CRAG Graph: configures workflow
src/workflow/graph.py - Runtime Settings → Document Scorer: configures model
src/evaluation/scorer.py - Runtime Settings → Hybrid Retriever: configures embedding
src/retrieval/hybrid.py