How ducbao210/finsight works
FinSight is a FinanceBench-focused, evidence-grounded financial QA system. Filings are parsed and chunked into indexes; a question enters the API and runs through corrective multi-hop retrieval, evidence grading, optional calculation, cited answer generation, and citation validation. A separate two-pass evaluation capability scores retrieval, citations, and generated answers. The README documents the major workflows; source excerpts confirm API initialization and dispatch plus key retrieval and graph components. Some unsampled implementation links are omitted where source evidence is unavailable.
Main entry points: runner.py Evaluation Runner · hybrid.py Hybrid Search · main.py FastAPI Service · nodes.py Reasoning Nodes
Components 24
Filing Ingestion
- Filing Parser
src/ingestion/parser.py - Filing Metadata
src/ingestion/metadata.py - Document Chunker
src/ingestion/chunker.py
Evidence Retrieval
- Vector Index
src/retrieval/dense.py - Lexical Index
src/retrieval/bm25.py - Dense Search
src/retrieval/dense.py - Hybrid Search
src/retrieval/hybrid.py - Optional Reranker
src/retrieval/reranker.py
Question Answering
- FastAPI Service
src/main.py - CRAG Orchestrator
src/graph/workflow.py - Workflow State
src/graph/state.py - Reasoning Nodes
src/graph/nodes.py - Reasoning Prompts
src/core/prompts/__init__.py - Financial Calculations
src/calculation/financial_metrics.py
Evaluation
- Evaluation Runner
src/evaluation/runner.py - Benchmark Dataset
src/evaluation/dataset.py - Retrieval Metrics
src/evaluation/retrieval_metrics.py - Citation Metrics
src/evaluation/citation_metrics.py - Generation Metrics
src/evaluation/generation_metrics.py - Evaluation Reports
src/evaluation/reporting.py
Shared Foundation
- Request and Result Schemas
src/schemas/__init__.py - Configuration
src/core/configs.py
Other components
- Questioner
- Benchmark Operator
Connections 23
- Questioner → FastAPI Service: submits question
README.md - FastAPI Service → CRAG Orchestrator: runs pipeline
src/main.py - CRAG Orchestrator → Reasoning Nodes: dispatches steps
src/graph/workflow.py - CRAG Orchestrator → Workflow State: uses state
src/graph/workflow.py - Reasoning Nodes → Reasoning Prompts: uses prompts
src/graph/nodes.py - Reasoning Nodes → Hybrid Search: retrieves evidence
src/graph/nodes.py - Hybrid Search → Dense Search: fuses search
src/retrieval/hybrid.py - Hybrid Search → Lexical Index: fuses search
src/retrieval/hybrid.py - Hybrid Search → Optional Reranker: optionally reranks
src/retrieval/hybrid.py - Dense Search → Vector Index: queries vectors
src/retrieval/dense.py - Reasoning Nodes → Financial Calculations: computes metrics
src/graph/nodes.py - FastAPI Service → Request and Result Schemas: validates responses
src/main.py - Filing Parser → Filing Metadata: extracts filing details
- Filing Metadata → Document Chunker: supplies metadata
- Document Chunker → Vector Index: indexes chunks
README.md - Document Chunker → Lexical Index: indexes chunks
README.md - Benchmark Operator → Evaluation Runner: runs evaluation
README.md - Evaluation Runner → Benchmark Dataset: loads questions
src/evaluation/runner.py - Evaluation Runner → Retrieval Metrics: scores retrieval
src/evaluation/runner.py - Evaluation Runner → Citation Metrics: scores citations
src/evaluation/runner.py - Evaluation Runner → Generation Metrics: scores answers
src/evaluation/runner.py - Evaluation Runner → Evaluation Reports: writes reports
src/evaluation/__init__.py - FastAPI Service → Configuration: loads settings
src/main.py