AutoIndex learns retrieval indexing programs and lifts Recall@100 by 8.4%
Sam O'Nuallain · hf · 2026-07-23
AutoIndex learns retrieval preprocessing as executable programs
AutoIndex treats document representation as an optimization target instead of a fixed preprocessing step. It searches over representation programs—executable transforms that can slice, enrich, normalize, reweight, or reorganize documents before they are indexed.
- The system uses validation-guided program search: agents inspect failures of the current index pipeline and synthesize candidate updates.
- Only updates that improve retrieval quality are kept.
- Evaluation on CRUMB, a benchmark of heterogeneous retrieval tasks, keeps BM25 fixed across experiments.
Results:
- +8.4% average Recall@100 over static full-document BM25
- +8.3% average nDCG@10
- Best task gains of +30.5% Recall@100 and +43.6% nDCG@10
The authors argue that document representation should be explicitly optimized, not hard-coded before retrieval starts. Code is available on GitHub.
Related event: AutoIndex Framework Proposes Optimization via Representation Programs(8 posts)→
More from coding & agent
- Dev torn on Cloudflare Agents SDK: full primitives but vendor lock-in — MikkoH · 2026-09-11
- Team-level AI agents: where should shared context and history live? — Al_Grigor · 2026-09-11
- Trust layer for money-moving AI agents: out-of-mandate actions can't get signed — Arpitbuilds · 2026-09-11
- Chaining dependent MCP tool calls: no rollback, duplicate risk — agentrsdg · 2026-09-11
- DeepMind-led paper makes design docs the source of truth, code disposable — SMART regenerates in 1.5-3h for ~$100 — Roger_M_Taylor · 2026-09-11
- Agent-built classifier labels 192k docs for $0.70 vs $13-26 with frontier LLMs — vanstriendaniel · 2026-09-11