Skip to main content

Stage 6: Summaries

The sixth stage is OKT's "system 2" thinking — a slow accumulation artifact that crystallizes knowledge per concept, accessible on demand. Summaries incrementally fold new facts into per-concept slices.

Entry point: (*SummarizeConceptsWorker).Work — backend/internal/taskmanager/tasks/summarize_concepts.go:118.

Per concept (summarizeOneConcept, summarize_concepts.go:185):

  1. Claim a per-concept lock: ClaimConceptForSummary sets concepts.summarizing_at; stale after 2 hours. This prevents two workers from summarizing the same concept concurrently.
  2. Load uncovered facts: ListUncoveredFactIDsForSummary — facts not in any existing summary's covered_fact_ids.
  3. Reconstruct the open slice's covered set + the new facts.
  4. Incremental slicing: produce complete slices of BatchSize (default 20) facts (is_complete=TRUE, frozen) plus one open remainder (is_complete=FALSE, keeps accumulating). The open slice is the one that grows as new facts arrive; complete slices are frozen and never touched again.
  5. One LLM call per slice (summarizer.Summarize) outside any tx.
  6. Upsert slice: CreateSummary or UpdateSummary.
  7. Chain out: enqueues synthesize_concept (summarize_concepts.go:395) only when at least one complete slice was written this pass. Open-slice-only updates wait for more facts.

The summarization prompt​

buildSystemPrompt — backend/internal/providers/summarization/prompt.go:33-101.

Key instructions:

  • "Reason ONLY from the provided facts. Do NOT use your training data."
  • Credulous: "Reflect what the facts say, not to judge them."
  • "PRIORITIZE VARIETY, NOT EXHAUSTION" — merge same-point facts, preserve distinct perspectives.
  • Inverted pyramid: core first, nuances last (survives truncation).
  • Strict word budget derived from MaxTokens.
  • Citation syntax: [text](<fact:fact_id>) — the fact: prefix discriminates fact citations from concept citations.
  • Radical source neutrality: no source gets prestige-based credibility.

buildUserMessage (line 111) renders each fact as N. [<fact:fact_id>] <text> (<attribution>).

The incremental design​

The slicing design is what makes summaries a slow accumulation artifact. As new facts arrive for a concept:

  • The open slice (is_complete=FALSE) accumulates them until it reaches BatchSize, then it freezes into a complete slice and a new open slice starts.
  • Complete slices are immutable — they represent a crystallized batch of knowledge.
  • A future reader can see the concept's knowledge grow over time by reading slices in order.

Key tables​

TablePurpose
okt_repository.concept_summariesPer-concept summary slices: sequence_num, is_complete, content, covered_fact_ids, model

The unique partial index uq_concept_summaries_concept_open on (concept_id) WHERE is_complete = FALSE ensures at most one open (accumulating) slice per concept.