feat: Added Tool Fetcher experiment and code.

[kevincogan]

Summary

Implemented the Tool Fetcher algorithm to let an agent dynamically discover and call tools via a single hub. Uses dense embeddings + optional Milvus vector index, with substring fallback. Reduces context overhead by fetching only relevant tools on-demand instead of loading all tools upfront.

What's included

• Algorithm: evaluator/algorithms/tool_fetcher_algorithm.py
  • Single exposed tool tool_hub that supports:
    • action='search' to retrieve relevant tools by name/description
    • action='call' to invoke a specific tool by exact name
  • Embedding-based retrieval using HuggingFaceEmbeddings with optional Milvus index; automatic fallback to substring match
  • Configurable hybrid search (dense + BM25), cross-encoder reranking, similarity thresholds, and indexed tool parts
  • Query-scoped active tool tracking and JSON responses for search/call actions
  • Tool usage logging via TOOL_LOG_PATH
• Config: evaluator/config/yaml/tool_fetcher_experiments.yaml
  • 17 total algorithms (16 variants + baseline) testing different embedding models, top_k values, indexed parts, and similarity metrics
  • Uses StableToolBench remote datasets (G1 category/instruction/tool) and reference answers
  • Includes tool selection, retrieval (Recall@k, AP), answer quality, and efficiency metric collectors

Quick test

# From repository root:
cd tool-rag-experiments

# Ensure .env file is configured with the correct values
# MILVUS_URL, QWEN_MODEL_URL, optional TOOL_LOG_PATH

python main.py -c evaluator/config/yaml/tool_fetcher_experiments.yaml