1.6 KiB
Extract single-word search terms from the user's query. These terms will be used as exact substring grep searches against markdown knowledge files.
Return two lists:
- exact_terms: specific identifiers, names, error codes — words that should appear literally in relevant files (single words, use root/stem forms for better grep matching)
- conceptual_keywords: broader topical terms that related files might contain (single words, prefer singular root forms over plurals for grep matching)
Keep each list to 3-8 terms. Every term must be a single word.
Important: These terms are used as grep filters, so choose terms with high discriminative power — terms that appear often in relevant files but rarely in unrelated files. Avoid overly generic terms like "code", "error", "function", "pattern", "file", "data", "config", "setup", "update" that would match most markdown files indiscriminately.
Examples:
Query: "How do I fix the flaky test in the bash permission classifier?" exact_terms: ["flaky", "bash", "classifier", "permission"] conceptual_keywords: ["retry", "reliability", "timeout", "intermittent"]
Query: "What's the pattern for adding a new MCP tool with streaming?" exact_terms: ["MCP", "streaming"] conceptual_keywords: ["register", "plugin", "handler", "protocol"]
Query: "Why does the ripgrep search timeout on large repos?" exact_terms: ["ripgrep", "timeout"] conceptual_keywords: ["scaling", "filesystem", "latency", "threshold"]