FinSight AI — multimodal agentic RAG over financial reports

A research assistant over financial reports (annual reports, investor decks) that understands text, tables, and charts, answers through a multi-tool LangGraph agent, exposes its capabilities over MCP so other MCP clients can use them, evaluates and monitors itself with an MLOps loop (MLflow + RAGAS + CI), supports users uploading their own documents with proper isolation, and ships a usage/quality dashboard for Power BI. Runs entirely free.

Why this project

Most portfolio RAG projects are "chat with a PDF" with no eval, no tests, and a static pre-loaded corpus. This one is built to answer the three questions that actually separate a demo from a production-minded project: does it stay correct over time (RAGAS + CI), is one user's data safe from another's (per-session vector isolation), and can anyone see how it's doing (the BI dashboard). See infra/aws/README.md for the deployment story.

Related MCP server: Nexla DocQA MCP Server

Architecture

Chat UI (Chainlit)
      │
      ▼
Agent orchestrator (LangGraph, tool-calling: search / chart-search / calc)
      │                                    │
      ▼                                    ▼
Multimodal RAG (Qdrant + sentence-       MCP server (same tools, exposed
transformers + Groq vision for charts)   to Claude Desktop / other clients)
      │
      ▼
AWS deployment (local by default; see infra/aws/README.md for the
S3 / Lambda / DynamoDB / Cognito production path)
      │
      ├──► MLOps loop: MLflow experiment tracking + RAGAS eval, run in CI
      └──► Telemetry → CSV export → Power BI dashboard

What's genuinely multimodal, agentic, etc. — and why

• Multimodal: chart/graph-heavy pages are rendered to an image and described by a free-tier vision LLM (Groq), then embedded as text alongside prose and tables — see src/ingestion/parser.py.

• Agentic: a LangGraph ReAct agent decides which tool to call (search_documents, extract_chart_data, calculate) and can chain multiple calls before answering — see src/agents/graph.py.

• RAG with real isolation: every vector is tagged source, user_id, session_id; retrieval always filters on these, so an uploaded document is never visible to another user — see src/rag/vector_store.py.

• MCP: the exact same tool functions are exposed as an MCP server, so Claude Desktop (or anything else speaking MCP) can use FinSight's retrieval — see src/mcp_server/server.py.

• MLOps: RAGAS scores (faithfulness, answer relevancy) are logged to MLflow on every CI run, not just eyeballed once — see src/eval/ragas_eval.py and .github/workflows/ci.yml.

• BI dashboard: every query and ingestion event is logged; export to CSV and open in Power BI Desktop (free) for a usage/quality dashboard — see dashboard/streamlit_ops.py for the quick in-app version.

Quickstart (100% free, no AWS account needed)

git clone <your-repo-url> && cd finsight-ai
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt

cp .env.example .env
# edit .env and paste a free Groq key from https://console.groq.com

# terminal 1
uvicorn src.api.main:app --reload

# terminal 2
chainlit run ui/chainlit_app.py -w

# terminal 3 (optional — usage dashboard)
streamlit run dashboard/streamlit_ops.py

Or with Docker: docker compose up --build.

Loading the shared corpus

Drop a few real annual-report PDFs into data/shared_corpus/ and run:

python -m scripts.ingest_shared_corpus   # see "what's next" below

Running the eval

python -m src.eval.ragas_eval
mlflow ui   # view the score trend at http://localhost:5000

Running tests

pytest -m "not slow"   # fast, no model download
pytest                 # full suite (downloads the embedding model)

What's next (honest roadmap, not finished-and-perfect)

• Split the single ReAct agent into explicit sub-agent graphs (retrieval specialist, numeric-reasoning specialist, writer) behind a supervisor node — the current single-agent version is simpler to debug and a reasonable place to start; multi-graph is the natural v2. Fill in src/eval/testset.json with real question/answer pairs once you've loaded actual filings — the placeholders there are structural, not real ground truth.

• AWS deployment per infra/aws/README.md.

Cost

Everything above runs for $0: Groq's free tier for LLM calls, local embedded Qdrant, local SQLite, GitHub Actions' free minutes for CI. The only signup required is a free Groq API key (no credit card).