A grounded orchestration framework for multi-agent collaboration on financial tasks leveraging the OSWorld environment, featuring Orchestrator-Worker dual-agent architecture with support for local database queries, GitHub integration, and financial data aggregation.
FinTalk.v is a multi-agent orchestration framework designed for financial tasks, featuring:
Grounded Cognition: Every piece of information is traceable to verifiable sources (database rows or deterministic computation results), systematically eliminating factual hallucinations.
Dual-Agent Architecture:
- Orchestrator Agent (Qwen3-8B): High-level reasoning, strategy planning, and answer synthesis
- Worker Agent (Qwen2.5-7B-Instruct-1M): Specialized skill execution with dynamic LoRA adapter serving
Worker Skill Set:
- Keyword Extraction (semantic parsing and entity recognition)
- Classification (intent classification and routing)
- NL2SQL (high-precision SQL generation)
MCP Integration Capabilities: Parallel Model Execution, Query Rewriting, Arbitration Mechanism, Rejection Detection, Correlation Analysis, Function Calling, Streaming NLG, Conversation Management
External Tool Integration: GitHub Search/Repo Manager, Google Search, Alpha Vantage (stock prices), NewsAPI (financial news)
Training Pipeline: SFT (NL2SQL LoRA with Qwen3-Embedding-8B vector semantic deduplication) + RL (verl framework + GRPO reinforcement learning fine-tuning)
Built-in Database: 607 companies, 2,970 management records, 2,208 shareholder records
Use Cases: Enterprise financial intelligence query and analysis, bank/financial institution internal knowledge base Q&A, multi-source financial data aggregation and comparison
Quick Start:
git clone https://github.com/boris-dotv/fintalk.v
cd fintalk.v
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt
python run.py
Environment Variables: GITHUB_TOKEN (required), GOOGLE_API_KEY, ALPHA_VANTAGE_KEY, NEWS_API_KEY, LLM_API_KEY (optional)