Skip to content

modules.md

Latest commit

 

History

History
49 lines (39 loc) · 3.17 KB

modules.md

File metadata and controls

49 lines (39 loc) · 3.17 KB
Raw

FinSight Modules

Languages: English | 中文

FinSight is organized into five modules. Each module has a clear responsibility and exchanges structured artifacts with the next stage.

Module Overview

Module Responsibility Main outputs Detailed docs
Frontend Browser interaction, /chat request flow, rendered answer cards, evidence display, and bilingual user experience. Chat UI state, user request, rendered answer, source cards. Local Frontend Chatbot
NLU and Retrieval Normalize the query, resolve entities, classify intent/topic/style, plan sources, retrieve documents and structured rows, rank evidence, and package traceable outputs. nlu_result, retrieval_result, coverage, warnings, debug trace. Query Intelligence
Numerical Analysis Convert structured market, fundamental, valuation, macro, and price-history rows into compact analytical signals. analysis_summary.market_signal, fundamental_signal, macro_signal, data_readiness. Numerical Analysis
Text Analysis Clean retrieved documents, detect language, filter entity-relevant sentences, and classify financial sentiment. SentimentResult, document-level SentimentItem, entity aggregates. Document Sentiment Analysis
LLM Summary and Prediction Generate frontend-ready answer JSON and follow-up question suggestions from compact evidence. answer_generation, next_question_prediction, citations, disclaimer. LLM Response Handoff

Data Flow

flowchart TD
  A["User query"] --> B["Frontend"]
  B --> C["NLU"]
  C --> D["Retrieval"]
  D --> E["Numerical Analysis"]
  D --> F["Text Analysis"]
  D --> G["LLM Summary and Prediction"]
  E --> G
  F --> G
  G --> H["Frontend answer"]
Loading

Boundaries

  • The frontend should not infer financial intent or source plans. It sends user input and renders returned evidence.
  • NLU and Retrieval should stay explainable and evidence-producing. They do not write final investment answers.
  • Numerical analysis summarizes structured evidence. It does not claim causal forecasts.
  • Text analysis classifies retrieved documents. It does not retrieve new sources independently.
  • LLM summary uses only compact evidence and cited evidence IDs. It must not invent missing facts.

Where to Change Things

Task Start here
Add a new endpoint or contract field query_intelligence/contracts.py, then update Query Intelligence.
Add a new source provider query_intelligence/integrations/, retrieval source planning, and provider config.
Add a new numerical indicator query_intelligence/retrieval/market_analyzer.py, then update Numerical Analysis.
Change frontend display query_intelligence/chatbot.py, then update Local Frontend Chatbot.
Change sentiment preprocessing or labels sentiment/, then update Document Sentiment Analysis.
Change LLM JSON output scripts/llm_response.py and /chat response mapping, then update LLM Response Handoff.