Es Knowledge Synthesis

Enterprise Search Knowledge Synthesis

When to use

Reach for this after a cross-source search returns overlapping, partial, or conflicting hits and the user needs an answer, not a link dump. You turn scattered snippets from chat, email, storage, and trackers into one sourced narrative with every claim traceable.

Not for: planning the search (es-search-strategy); running the fan-out (es-search); or a single unambiguous hit that needs no reconciliation — just quote it with its source.

Method

1. Cluster by claim. Group hits around the distinct facts they assert, not by which source they came from.
2. Deduplicate. Collapse restatements of the same fact into one claim, but keep every distinct source pointer so nothing loses provenance.
3. Reconcile conflicts. Decision: when sources disagree, prefer the most recent and most authoritative — but surface the disagreement explicitly rather than silently picking a winner.
4. Attribute every claim to its source(s): who, where, when. Decision: if a statement has no source, it does not enter the answer — flag it as a gap instead.
5. Rate confidence per claim: High (multiple independent recent sources agree), Medium (single credible source), Low (stale, indirect, or contested).
6. Fill or flag gaps. State what the sources do not cover instead of interpolating, and point to where the missing piece likely lives.

Example

Three hits on "current API rate limit": a wiki page (6 months old) says 100/min, a Slack thread (last week) from the platform lead says "we raised it to 300/min," a ticket (last week) marked done reads "bump rate limit to 300." You cluster into one claim "rate limit = 300/min," mark it High (two recent independent sources), note the wiki's 100/min as a stale conflict that was superseded, and flag that the wiki page needs updating.

Pitfalls

• Averaging or silently choosing between conflicting numbers instead of showing both and the reason one wins.
• Dropping a source pointer during dedupe, so a claim becomes unverifiable.
• Rating confidence High off a single source, or Low purely because a source is old when it is uncontested.
• Filling a coverage gap with a plausible guess rather than naming it as unknown.

Output format

Answer: <2-5 sentences of the settled facts>

Supporting claims:
- <claim> — [source, author, YYYY-MM-DD] — High|Medium|Low
- ...

Conflicts:
- <topic>: <side A> vs <side B> → preferred <X> because <reason>

Confidence overall: <High|Medium|Low> — <driver>
Gaps / next lookups: <what's unknown> → check <where>