-
fix(langgraph): default search.fetch.mode to summary_focus_query (#3793)
frostbyte_neo 发布于
2026-05-02 13:01:35 +00:00 The 'full' default introduced in PR #3680 (v1.6.3) floods small/mid
local-model contexts with full extracted page text, which both slows
each example down (~60% in a SimpleQA run with qwen3.6:27b) and
degrades accuracy (95.7% → ~80% on a 25-example sample, with two
queries hitting the iteration limit and several returning 0 results).Switch the default to 'summary_focus_query' so the LLM extracts only
spans relevant to the agent's per-fetch focus question AND the original
research query — the latter disambiguates a vague focus and matches the
designed-for behaviour described in the original setting's help text.Migration 0009 flips existing rows whose value is exactly 'full'; users
who explicitly chose summary_focus, summary_focus_query, or disabled
are left untouched. Users who deliberately picked 'full' can re-select
it from the settings UI.下载附件