Skip to content

Materialization

Why Occam returns compiled content instead of raw HTML.

Status: STABLE (Markdown + budgets) · some structured/diff/cache paths LIMITED or EXPERIMENTAL

The point

Agents have finite context windows. Occam’s job after acquisition is to produce usable content:

  • primary output: Markdown
  • sized to a token budget
  • optionally focused on what you asked for
  • optionally emitted as structured blocks, tables, feeds, or chunks
  • optionally compared as a diff / conditional re-check

The hash on a receipt binds this compiled form — not necessarily the origin’s raw bytes.

Token budgets

Mechanism Role
max_tokens Explicit budget for this call
occam_client_capabilities / OCCAM_CLIENT_CONTEXT_TOKENS Ambient budget (~20% of declared context, clamped)
fit_markdown + focus_query Prune toward relevance even when the full page fits; retain locally coupled instruction blocks before budgeting

Omit knobs for a first read — defaults are intended to be enough.

Full advantage + knob flashcard (agents and humans): Why Occam. There is no public MCP parameter to select an alternate knowledge codec; live output is Markdown via the default passthrough codec.

Structured outputs (opt-in)

On occam_transcode and related tools, parameters such as json_blocks, json_tables, json_feed, and chunk modes add sidecars. They are not required for ordinary reading.

Omitted content may be summarized in an omission manifest when fitting drops material — honesty about what was cut.

Focused fitting keeps locally coupled commands, procedural lists, and adjacent explanations together. Independent navigation links remain individually filterable. When a selected instruction cannot fit, an explicit omission replaces it rather than a heading or introduction without its steps.

compile.omitted can report focus_filtered even when compile.truncated is false: filtering and budget truncation are different operations. SNIP comments include removals within retained sections when space permits. completeness.incompleteReason=focus_body_filtered means the selected answer was lost during filtering; increasing the budget is not a remedy. focus_body_truncated denotes budget loss, with an estimated suggestedMinTokens above the current budget. Completeness concerns the selected answer, not every section of the source.

Digest defaults to fit_markdown=true; transcode defaults to false. Digest's focusMatched measures lexical overlap in the excerpt, while focus.status measures structural section matching. focusMatched:true can coexist with focus.status:weak and does not prove answer completeness.

Cache and identity

  • Default: live extract every call (no file cache).
  • Opt-in: cache_ttl_s > 0 enables a local response replay (not a CDN).
  • Cache / materialization keys include focus-sensitive identity (including URL fragment where applicable).
  • if_none_match / diff_against compare against Occam’s content hash of compiled output.