IbrahimKhan7208/investment-research-router
The IbrahimKhan7208/investment-research-router is a 1.5 billion parameter Qwen2.5-1.5B-Instruct model, fine-tuned through a CPT, SFT, and GRPO pipeline. It is designed to replace multiple LLM calls in a LangGraph-based investment research agent with a single structured-output call. This model excels at routing investment research questions to appropriate tools (RAG, WEB, STOCK) and extracting structured filters like companies and years, generalizing beyond hardcoded entity lists.
Loading preview...
Investment Research Router Model
This model is a specialized Qwen2.5-1.5B-Instruct (1.5B parameters) designed to streamline multi-tool investment research agents. It replaces a two-call, 120B-class LLM system with a single, structured-output call, significantly improving efficiency and capability.
Key Capabilities
- Unified Tool Routing & Extraction: Collapses
classifierNodeandextractFiltersFromQuestionsinto one call, providing tool routing and structured data extraction (companies, years, search query) for RAG, WEB, and STOCK tools. - Generalized Entity Extraction: Extracts companies and years without relying on hardcoded lists, allowing for research on any named company.
- Robust Training Pipeline: Developed using a full CPT → SFT → GRPO pipeline, adapting domain vocabulary, teaching the routing/extraction task, and refining for edge cases.
- Improved Accuracy: Achieves 96.8% tool-routing accuracy and 94.1% RAG field accuracy on in-scope questions, significantly outperforming the 120B baseline.
- Efficiency: Reduces average LLM calls per query by 33-38% compared to the original two-call system.
Good for
- Developers building LangGraph-based investment research agents needing efficient, structured tool routing.
- Applications requiring dynamic extraction of investment-related entities (companies, years) from natural language queries.
- Scenarios where a smaller, specialized model can replace larger, more general LLMs for specific, high-volume tasks.