Community Summarization
For "global" questions about a person, project, or theme, naive retrieval pulls a handful of edges and misses the gestalt. The right answer is the cluster summary — the writeup of an entity community produced during consolidation — not raw turns.
When to Use
- The user asks an open question about a person or topic ("how have things been with Sara?").
- A retrospective or check-in across a long period.
- Drafting an introduction, reference, or thank-you that needs more than one anecdote.
- The query touches many small facts that are individually trivial but cumulatively meaningful.
Workflow
- Detect the global shape. Pronouns, possessives, theme words ("the project," "your year"). If retrieval would need 50+ rows to cover, you want a summary.
- Look for the precomputed community summary first. AutoDream produces them during idle cycles; they're cheaper than synthesizing on the hot path.
- If no summary exists, run a one-shot consolidation. Score the top-K most-recalled, highest-weighted edges in the cluster and summarize. Cache the result.
- Cite, even at the cluster level. A summary without provenance becomes hallucination on revisit.
- Refresh when stale. Summaries older than a month or behind a wave of new edges should be invalidated.
Anti-patterns
- Summarizing on every query (latency cost).
- Summarizing without writing back the summary (next query repeats the work).
- Letting one outlier turn dominate the summary because it was emotionally loaded.
Examples
Example 1: "How are things with the design team?" → Pull the design-team community: members, recent decisions, mood arc, open threads. One paragraph, ~300 words, with two or three concrete anchors.
Example 2: "Tell me about my mom's hobbies." → Cluster is small but specific; precomputed summary exists. Read directly, augment with anything new since last_summarized.
1---2name: community-summarization3description: Community Summarization4---5# Community Summarization67For "global" questions about a person, project, or theme, naive retrieval pulls a handful of edges and misses the gestalt. The right answer is the cluster summary — the writeup of an entity community produced during consolidation — not raw turns.89## When to Use10- The user asks an open question about a person or topic ("how have things been with Sara?").11- A retrospective or check-in across a long period.12- Drafting an introduction, reference, or thank-you that needs more than one anecdote.13- The query touches many small facts that are individually trivial but cumulatively meaningful.1415## Workflow161. **Detect the global shape.** Pronouns, possessives, theme words ("the project," "your year"). If retrieval would need 50+ rows to cover, you want a summary.172. **Look for the precomputed community summary first.** AutoDream produces them during idle cycles; they're cheaper than synthesizing on the hot path.183. **If no summary exists, run a one-shot consolidation.** Score the top-K most-recalled, highest-weighted edges in the cluster and summarize. Cache the result.194. **Cite, even at the cluster level.** A summary without provenance becomes hallucination on revisit.205. **Refresh when stale.** Summaries older than a month or behind a wave of new edges should be invalidated.2122## Anti-patterns23- Summarizing on every query (latency cost).24- Summarizing without writing back the summary (next query repeats the work).25- Letting one outlier turn dominate the summary because it was emotionally loaded.2627## Examples28**Example 1:** "How are things with the design team?" → Pull the design-team community: members, recent decisions, mood arc, open threads. One paragraph, ~300 words, with two or three concrete anchors.2930**Example 2:** "Tell me about my mom's hobbies." → Cluster is small but specific; precomputed summary exists. Read directly, augment with anything new since `last_summarized`.
sethdford