Skip to content

Deep Dive: Dynamic Routing (BM25)

In SILO V2, an Agent doesn't have to know specific tool names upfront. Instead, it uses Semantic Discovery to find the right tool for the job.

1. How Search Works

When you call silo_search(query="get market data"), SILO performs a hybrid search across all installed skills in the hub:

  1. Exact Match: Tools or namespaces that exactly match the query (or are substrings) are ranked highest.
  2. BM25 Fallback: If no exact matches are found, SILO uses the BM25Okapi algorithm (Best Matching 25) to rank tool descriptions and instructions semantically.
  3. Cross-Namespace: Search is not limited to a single skill; it scans the entire hub.

2. Optimizing for Agents

To make your tools easily discoverable, follow these best practices in your skill.py:

  • Descriptive Names: Use get_stock_price instead of fetch_data.
  • Rich Docstrings: The text in your function's docstring is indexed by the search engine.
    • Good: """Fetches the current trading price for a stock ticker symbols like AAPL."""
    • Bad: """Gets price."""
  • Skill Instructions: Use @skill.instructions() to describe the overall capability of the skill (e.g., "This skill is the source of truth for all financial data").

3. The silo_search Tool

The MCP server exposes silo_search. Here is how an Agent typically interacts with it:

Agent: "I need to check the company's annual revenue. Let me search for an appropriate tool."

(Agent calls silo_search(query="annual revenue"))

SILO: Returns fin-ops:get_financials with its description and schema.

Next: See how SILO handles Interactive Approvals.