Skip to content

Chapter 13 — Typed field extraction

Part D — Site-specific · Prerequisites: Ch 11, Ch 2 · Next: Chapter 14

Mental model

A separate spine with its own worker and narrower failure taxonomy.

Recipe D: occam_playbook_resolve → schema match → occam_extract_knowledge. Caller must supply knowledge_schema (or rely on resolved schema). There is no schema-free mode. Output is facts[] plus Receipt telemetry—not Receipt v1.

Hard dependency: PS-4 extraction requires PS-5 playbook resolution.

Explanation

What the tool returns

  • facts[] — Typed fields per caller schema (selectors, types, row shapes where supported).
  • meta.koId — Knowledge object identifier when configured.
  • Receipt object { confidence, elapsedMs } — Extraction telemetry only (OD-5). Unsigned. No contentHash, no signature, no Merkle root. occam_verify does not accept it.
  • confidence in telemetry is heuristic—often 0.0; not a correctness guarantee.

What this path skips

Compared to occam_transcode (Chapter 4):

  • No OccamRouter post-processor pipeline (thin/challenge reclassification path differs)
  • No token budget (max_tokens ignored)
  • No Receipt v1 signing
  • Browser leg may spawn throwaway Playwright per call (no pool reuse on browser fallback)
  • session_profile storageState is applied on the CSS extract browser-fallback leg (same as transcode).

Row mode

Host parsers do not set base_selector for row mode—row mode is dead. Document field/list selectors only.

Security (do not point at untrusted URLs)

  • css-extract lacks DNS pinning and body cap on some paths.
  • Nuxt readNuxtPath evaluates page-controlled state via eval().

Typed extraction is for known site shapes with operator-authored schemas, not arbitrary open-web scraping.

Task R step 10

Extract {limit, window, scope} via schema saved in Chapter 12 instead of re-parsing markdown.

CHECK

NETWORK

  1. Call occam_extract_knowledge and occam_transcode on the same URL with max_tokens set on both.
  2. Assert: extract ignores budget; transcode respects it.
  3. Assert: extract Receipt.confidence is heuristic (often 0.0); no signed receipt.signed envelope on extract response.

Attempt occam_verify on extract telemetry—it must reject/not apply.

Common misconception

"occam_extract_knowledge returns a Receipt, so its output is verifiable."

The wire field named Receipt is extraction telemetry, not Occam Receipt v1. Only signed extraction envelopes from transcode/claim paths verify.

Limitations

  • Worker timeout hardcoded ~45s.
  • No Merkle citations from extract path—for quotable proofs, use transcode + claim_check (Chapter 14, future Ch 16).
  • Values are as-extracted, unsigned—schema validation is structural, not truth.
  • Do not present as safe for untrusted URLs until EF-013/043 mitigations ship.

Public docs: Structured extraction · Examples: structured extraction · Tools: occam_extract_knowledge

Next chapter: Chapter 14 — What a receipt proves