AI-Powered Financial Research Platform — From data to insights, fully automated.
QuantHarbor is an end-to-end financial intelligence platform that transforms unstructured market documents into citation-grounded insights through automated retrieval, analysis, and report generation. It combines a multi-agent research system with a document-to-insight RAG pipeline, all wrapped in an interactive web demo UI.
Landing Page — Professional entry point with feature highlights
Main Interface — System configuration, task management, and execution monitoring
- Screenshots
- Key Features
- Quick Start
- Architecture
- Project Achievements
- Advanced Usage
- License
- Acknowledgments
-
Document-to-Insight RAG Pipeline Ingests large PDF corpora, builds searchable vector indexes, and produces citation-grounded research output with file/page provenance and similarity scores.
-
Multi-Agent Financial Research Four specialized agents (Data Collector, Data Analyzer, Report Generator, Deep Search) collaborate through a shared variable space using the Code Agent with Variable Memory (CAVM) architecture.
-
Professional Report Generation One-click generation of 20,000+ word financial reports with automated charting. Reports are exported as Markdown, DOCX, and PDF with publication-grade formatting via Pandoc.
-
Automated Charting with VLM Feedback Built-in vision agents automatically correct chart issues (missing legends, wrong scales, low information density) through iterative visual feedback loops.
-
Deep Research with Evidence Tracing Every conclusion is derived from a transparent Chain-of-Analysis, with strict citations linking back to original data sources for verifiable insights.
-
Provider Flexibility Supports OpenAI and DeepSeek with configurable primary/fallback routing. Compatible with aggregator endpoints like OpenRouter.
-
Interactive Web Demo UI Full-stack web application with React frontend (Ant Design) and FastAPI backend. Includes system configuration, task management, real-time execution monitoring via WebSocket, and report browsing/download.
-
Checkpoint & Resume Long-running research tasks can be paused and resumed from the last checkpoint, with full agent state preservation.
-
Reproducible Local Artifacts Saves parsed corpus and index artifacts for deterministic reruns and debugging.
- Python 3.10+
- Pandoc (for polished DOCX/PDF export)
- Node.js (optional, for the web UI)
- API keys for your LLM stack (LLM, VLM, Embedding, Search)
# Clone the repository
git clone https://github.com/YOUR_USERNAME/quant_integrated_project.git
cd quant_integrated_project
# Create virtual environment
python3 -m venv .venv
source .venv/bin/activate
pip install -U pip
# Install Python dependencies
pip install -r requirements.txtInstall Pandoc (recommended):
# macOS
brew install pandoc
# Linux
sudo apt-get install pandocBuild the web UI (optional):
cd demo/frontend
npm install
npm run buildQuantHarbor uses a two-layer configuration:
.env— model endpoints & API keys
cp .env.example .env
# Fill in your API keys and model endpointsmy_config.yaml— research target & tasks
target_name: "Apple Inc."
stock_code: "AAPL"
target_type: "financial_company" # financial_company | macro | industry | general
output_dir: "./outputs/my-research"
language: "en" # en or zhCLI (full pipeline)
python run_report.pyWeb Demo
# Backend
cd demo/backend && python app.py
# Frontend (in a separate terminal)
cd demo/frontend && npm run devRAG Pipeline (document-to-insight)
# 1) Provider smoke test
python3 scripts/smoke_test_providers.py --provider auto
# 2) Ingest PDF corpus
python3 scripts/ingest_pdfs.py --source-dir "./data/raw_pdfs" --recursive true
# 3) Build retrieval index
python3 scripts/build_index.py
# 4) Generate report output
python3 scripts/run_pipeline.py --question "Summarize investment risks and opportunities."
QuantHarbor combines two complementary pipelines:
A multi-stage, memory-centric pipeline: Data Collection, Analysis + VLM chart refinement, Report drafting & polishing, Rendering. Each agent runs in a shared variable space with resumable checkpoints.
| Agent | Purpose | Key Inputs | Outputs |
|---|---|---|---|
| Data Collector | Route and gather structured/unstructured data | Task, ticker/market | Normalized datasets in memory |
| Deep Search Agent | Multi-hop web search + content fetch with source validation | Task, query | Search snippets + crawled pages with citations |
| Data Analyzer | Code-first analysis, charting, VLM critique | Task, collected data | Analysis report, charts + captions |
| Report Generator | Outline, sections, polish, cover/reference, DOCX/PDF | Task, analysis/memory | Publication-ready report (MD/DOCX/PDF) |
A modular document-to-insight pipeline:
- Ingestion — Parse PDF pages and detect low-extraction pages for OCR follow-up
- Indexing — Chunk text and build embedding index
- Retrieval — Retrieve top-k relevant chunks per query via cosine similarity
- Reasoning — Analyze evidence with LLMs under provider routing
- Report Generation — Return concise, citation-grounded research output
quant_integrated_project/
├── src/ # Multi-agent system (CAVM architecture)
│ ├── agents/ # Data Collector, Analyzer, Report Generator, Search Agent
│ ├── config/ # Configuration management
│ ├── template/ # Report outline templates
│ ├── tools/ # Financial, macro, industry, web tools
│ └── utils/ # LLM interface, logging, prompt loader
├── src_rag/ # RAG pipeline modules
│ ├── agents/ # Planner, Retriever, Analyst, Writer agents
│ ├── core/ # LLM client, embeddings, vector index, pipeline
│ └── tools/ # PDF ingest, web search
├── demo/ # Full-stack web demo
│ ├── backend/ # FastAPI app with WebSocket logs
│ └── frontend/ # React SPA (Ant Design)
├── scripts/ # CLI pipeline scripts
├── assets/ # Images, videos, example reports
├── data/ # PDF corpus and parsed pages
├── artifacts/ # Vector index (embeddings + metadata)
├── config/ # RAG pipeline configuration
└── docs/ # Advanced usage documentation
- Processed 138 PDFs end-to-end in one corpus ingestion run.
- Extracted 4,577 pages into normalized machine-readable records.
- Built a searchable vector index with 8,309 chunks.
- Generated structured, citation-grounded outputs with file/page provenance.
- Reduced initial research synthesis time from manual multi-document review to minutes-level automated generation for first-draft outputs.
- Enabled deterministic reruns via persisted parsed/index artifacts, improving reproducibility for iterative analysis.
- Introduced provider failover routing (OpenAI + DeepSeek) to improve runtime robustness under API availability constraints.
See docs/ADVANCED_USAGE.md for comprehensive technical documentation including:
- API keys & model configuration
- YAML config file reference
- Prompt system & customization
- Custom outlines & report templates
- Chart styling & color palettes
- Adding custom tools and agents
- Checkpoint & resume system
- Complete code examples
- Parsed pages:
data/parsed/pages.json - Index metadata:
artifacts/rag_index/metadata.json - Index embeddings:
artifacts/rag_index/embeddings.npy - Generated reports:
outputs/<target_name>/(MD, DOCX, PDF) - Agent logs:
outputs/<target_name>/logs/
This project is licensed under the GNU General Public License v3.0. See the LICENSE file for details.
- Advanced usage: docs/ADVANCED_USAGE.md
- API setup: docs/api_setup_openai_deepseek.md
- OCR and extraction notes: docs/pdf_coverage_and_ocr.md
- Migration notes: docs/migration_notes.md
- AkShare — Financial data APIs.
- eFinance — Stock data.
- Crawl4AI — Web crawling.
- Ant Design — React UI component library.