Context as Infrastructure: knowledge compilation for agentic retrieval

Source: arXiv:2610.41172 (RAGFlow team, InfiniFlow), surfaced by DAIR.AI on X (thread) · raw: raw/twitter/feed/2026-11-02-evening-230005-ranked.json

TL;DR

A search harness with only BM25 and grep on MiniMax M3 answers 85% of BrowseComp-Plus and 91% of FRAMES without vector or compiled indexes. Accuracy is largely solved; context cost is not, because the loop repeats queries and rereads documents. Knowledge compilation (summary trees, PageIndex-style outlines, wikis, graphs) gave no detectable FRAMES accuracy gain, and the outline degraded at scale. Its value is as a supply layer: a navigation tree with claim-evidence leaves cuts query tokens 91% when the leaf suffices (~43% of queries) and adds 1.9% when it does not, a 38% average reduction with a 2% break-even sufficiency rate. Savings cost retrieval rounds (tokens halved, latency ~3x on one subset); a decision gate cut latency from 140 s to 30 s.

Compile the knowledge once, hand the agent the smallest sufficient set

The tree saves tokens only when its claim-evidence leaf is enough; a gate caps the latency cost.

flowchart LR
  D["Documents<br/><small>raw corpus</small>"] --> K["Compiler<br/><small>LLM builds tree</small>"]
  K --> T["Claim-evidence leaves<br/><small>pre-digested facts</small>"]
  Q["Query<br/><small>agent loop</small>"] --> G{"Gate<br/><small>is the leaf enough?</small>"}
  T --> G
  G -->|yes, ~43%| A["Answer<br/><small>91% fewer tokens</small>"]
  G -->|no| B["BM25 + grep<br/><small>full retrieval</small>"]
  B --> A
  classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
  classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
  classDef loop fill:#fff3bf,stroke:#f08c00,color:#1b1b1b,stroke-width:2px
  classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
  class D,Q input
  class K,T core
  class G,B loop
  class A exit
  linkStyle 4 stroke:#2f9e44,stroke-width:2px

Blue is input, purple is the offline compilation, amber is the online decision and fallback, green is the answer.

Relation to prior wiki pages

Extends agent harness engineering: another harness result where measurement in tokens and latency matters more than accuracy, following DeepSeek Harness's reported 87.9% token cut (Tsinghua and Microsoft, circulating 11-02).
Same lesson as the NVIDIA KV-cache infra post (11-02, summary): savings only count in the unit the bill uses.
Relevant to agent memory: ontology, timeline and memory artifacts pay off in governance and continuity that benchmarks do not see.
Counterpoint: Mixedbread (AI Engineer talk, 11-02) reports 36% to 98% accuracy from better multimodal late-interaction search on PDFs, where the OCR-chunk-embed stack loses information before the harness sees it.