GraphRAG Finance

A financial analysis system that combines Knowledge Graphs, Vector RAG, and Deterministic Calculations to answer questions about company financial reports.

Built from scratch � no LangChain, no LlamaIndex. Just Python, Neo4j, ChromaDB, and a local LLM (Mistral via Ollama).

---

What It Does

Upload a financial PDF (10-K annual report) and ask questions like:

• "What risks does Apple face in China?"
• "Who is the CEO?"
• "What is the gross margin?"
• "Compare Apple and Microsoft's risk profiles."

Every answer is sourced from three systems working together:

Source	Used for
Knowledge Graph (Neo4j)	Relationships, risks, people, connections
Vector RAG (ChromaDB)	Specific facts, quotes, narrative text
Deterministic Engine	Exact financial numbers (revenue, margins)

---

Architecture

PDF
 +-- ChromaDB (vector store)  ---- Vector RAG path
 +-- financial_metrics.json   ---- Deterministic path
 +-- Neo4j (knowledge graph)  ---- Graph RAG path
          ?
    hybrid_ask() � combines all three ? LLM ? answer

---

Requirements

• Python 3.10+
• Ollama with Mistral pulled (ollama pull mistral)
• Neo4j Desktop with a local database running on bolt://localhost:7687

---

Setup

# 1. Clone the repo
git clone https://github.com/sudhendrak04/GraphRAG-finance.git
cd GraphRAG-finance

# 2. Create and activate a virtual environment
python -m venv .venv
.venv\Scripts\activate        # Windows
# source .venv/bin/activate   # Mac/Linux

# 3. Install dependencies
pip install -r requirements.txt

# 4. Make sure Ollama is running with Mistral
ollama pull mistral
ollama serve

---

Data

Place your company PDFs in the data/ folder like this:

data/
  Apple/
    Apple23 report.pdf
  Microsoft/
    microsoft_23'-10k.pdf

---

Running

Step 1 � Build the vector store and extract metrics

Open code/evaluation.ipynb and run run_full_pipeline() for each company.

Step 2 � Build the knowledge graph

Open code/build_company_graph.ipynb and run build_company_graph() for each company.
This sends every text chunk to Mistral and stores extracted entities/relationships in Neo4j.
Note: This takes 15-30 minutes per company depending on PDF size.

Step 3 � Start the web UI

python app.py

Open http://127.0.0.1:5000 in your browser.

---

Project Structure

GraphRAG-finance/
+-- app.py                    Flask backend (3 API routes)
+-- pipeline.py               PDF ingestion, ChromaDB, metrics extraction
+-- requirements.txt
+-- src/
�   +-- answer_engine.py      Deterministic math + Vector RAG
�   +-- evaluator.py          Accuracy testing
�   +-- graph_store.py        Neo4j read/write
�   +-- graph_extractor.py    LLM-based entity/relationship extraction
�   +-- graph_pipeline.py     Orchestrates graph building from chunks
�   +-- graph_ask.py          Graph traversal question answering
�   +-- hybrid_ask.py         Combines all three sources
�   +-- compare_ask.py        Side-by-side company comparison
+-- templates/
�   +-- index.html            Web UI
+-- data/                     Place your PDFs here

---

Neo4j Schema

Nodes:   (:Entity {name, type, company})
Edges:   (:Entity)-[:RELATES_TO {type, company}]->(:Entity)

Entity types: Company, Person, Product, Risk, Location, Regulation, Event

---

Limitations

• Graph extraction quality depends on Mistral (7B). Larger models give better results.
• Hub nodes (the main company node) have 100+ connections � retrieval is capped at 60 paths per query.
• Graph building is slow locally (~3-5 sec per chunk). Using an API-based model (GPT-4o) reduces this to ~30 min total.

---

Tech Stack

Component	Tool
PDF parsing	PyMuPDF
Vector store	ChromaDB
Embeddings	sentence-transformers (all-MiniLM-L6-v2)
Knowledge graph	Neo4j
LLM	Mistral 7B via Ollama
Web framework	Flask