Skip to content

Chapter 10 — Many sources in one call: digest

Part C — Breadth · Prerequisites: Ch 7, Ch 9 · Next: Chapter 11


Mental model

Digest is the acquisition spine run N times under one budget—plus combine ordering hints. It is not a discovery tool and not a synthesizer. suggestedReadOrder is a read hint, not generated summary across sources.

Rule: several URLs → one occam_digest, not N× occam_transcode—when one combined budget and focus apply.


Explanation

What digest does

occam_digest(urls, …):

  1. Validates the URL list (clamps apply—silent truncation at limits).
  2. Applies SSRF/session preflight to the whole batch (not always per-item isolated).
  3. For each URL, runs acquisition + compile similar to transcode (per-item policies possible).
  4. Applies one shared focus/budget across items (focus_query, global max_tokens, per_url_max_tokens).
  5. Returns per-item results with individual ok, markdown snippets, failures, and reduced receipts.

What digest gives up vs N transcodes

Aspect Digest N transcodes
Per-item full sidecars Limited / absent on items Full transcode sidecars each call
Playbook overlay on items Not on digest items Available per transcode
Receipt strength Reduced — content hash only, no block leaves / time anchor Full Receipt v1 when enabled
Failure isolation Batch preflight can affect whole call Independent per URL
Parallelism Internal fan-out, one in-flight per host (cross-host up to cap), one MCP call Agent-orchestrated

Per-item failures

Digest can return ok:true overall with some items ok:false—each item failure is typed. Task R: five reference pages + one deliberate 404 in the batch.

Honest framing

Digest concatenates and orders bounded per-URL materializations under shared focus—it does not merge facts, resolve contradictions, or judge truth.


CHECK

NETWORK

  1. occam_digest five URLs including one known 404.
  2. Assert: overall response succeeds structurally with a typed per-item failure for the 404.
  3. Assert: successful items' receipts lack full block-leaf structure (reduced receipt).

Compare token totals vs five separate transcodes with the same per_url_max_tokens—digest should win on orchestration overhead, not on magic compression.


Common misconception

"Digest summarizes across sources."

No model synthesizes across URLs. You receive bounded excerpts and hints; synthesis remains the agent's job—with citation discipline from each item's hash/receipt where present.


Limitations

  • Reduced receipts weaken per-item trust vs transcode (Chapter 14).
  • Batch SSRF/session preflight is whole-batch—not per-item session granularity.
  • Silent URL count/token clamps—verify list length in response.
  • Same-host URLs are extracted sequentially (polite); only distinct hosts fan out in parallel.
  • For claim-level proofs, you may still need full transcode receipts on critical URLs.

Public docs: Research multiple · Examples: research several · Tools reference (occam_digest)

Next chapter: Chapter 11 — Playbooks in band