Skip to content
AutoResearch
StaleDiscardedLow bandKnowledge Base

KB chunk-size sweep for retrieval precision

Baseline
74%
Final
70%
Delta
-4 pts
Variants
5
Objective

What we set out to improve

Raise retrieval precision on school and camp source documents by tuning the KB chunk size.

Discarded

Discarded. No alternative chunk size beat the 512-token baseline; smaller and larger chunks both regressed precision. The baseline configuration was retained unchanged.

Iterations

Variants we tried

Each variant and its coarse objective metric. The kept variant is marked; bars are relative to the best run.

  • 1Baseline — 512-token chunksWinnerLow74%
  • 2Variant A — 256-token chunksLow70%
  • 3Variant B — 384-token chunksLow72%
  • 4Variant C — 768-token chunksLow69%
  • 5Variant D — 1024-token chunksLow66%
Run

Stages

  1. baseline

    Succeeded · 1.8s

  2. variant run

    Succeeded · 7.2s

  3. eval

    Succeeded · 1.1s

Output

Artifacts and what shipped

Redaction-safe artifact previews, diffs, metric tables, and prompt variants with sensitive text removed.

  • Metric table

    Precision by chunk size (baseline stays best)

  • Report

    Discarded: no variant beat the 512-token baseline

What you can see, and what is hidden

Every projection on this page is redaction-safe by construction. Redaction level: Sample content, curated, public-safe excerpts only.

Shown

  • Identifiers & counts
  • Closed-enum statuses
  • Coarse quality / resource bands
  • Timestamps & freshness

Intentionally hidden

  • Raw prompts
  • Raw documents
  • raw tool log
  • Raw trace spans
  • Embedding vectors
  • Free-text feedback
  • Auth internals & secrets
  • Secrets

Related in the Lab

All experimentsThe full gallery, grouped by product area.Family Manager observatoryThe capture → approve → route loop.News DigestRanked items, source health, and feedback.Knowledge Base observatoryWhere promoted learnings are written.