As a developer, I want to add BM25 full-text search before agentic search to reduce usage cost

[kiyotis]

Situation

The current agentic search invokes the LLM for every query, which incurs significant Claude API costs (primarily cache writes). BM25 full-text search can pre-filter candidates at zero LLM cost.

The existing keyword-search workflow is confusingly named and overlaps in purpose — it will be removed.

Pain

Developers operating nabledge face high per-query costs because every search invokes the LLM even when keyword-based filtering could eliminate most irrelevant pages first.

Benefit

• Developers can run nabledge at lower cost per query
• The codebase becomes clearer by removing the misleading keyword-search workflow

Input Documents

• Cost Optimization Report ② Current Architecture (Agentic Search)
• Cost Optimization Report ① Architecture Overview

Success Criteria

• Implementation is based on the input documents (changes allowed during design/implementation)
• v6 benchmark answer accuracy is not degraded compared to the previous run
• v6 benchmark cost and execution time are lower than the previous run
• The mechanism validated on v6 is rolled out to all other versions (v5, v1.4, v1.3, v1.2)
• keyword-search workflow is removed from all versions

🤖 Generated with Claude Code