battam1111/omniseek/omniseek-investigate

/omniseek-investigate

Deep investigation METHODOLOGY for OmniSeek. Teaches judgment patterns for composing retrieval primitives into multi-step evidence gathering. Tool facts (signals, handles, _meta, evidence graph schema, walled source mechanics) are in the server instructions, not here. This skill teaches HOW to use those facts well.

1. The 3-turn rhythm (sweep, zoom, structure)

Every deep investigation follows the same beat:

WAVE 1 (sweep): fire several independent tools in parallel via omniseek_gather. A broad omniseek_search plus any tools whose results you need to make your first judgment call (e.g. omniseek_resolve_identity for a person question, omniseek_field_skeleton for a field question). One round-trip, all results at once.

Judge: read Phase A signals, handles, and _meta (per server instructions sections 3-5). Decide what to zoom on: which walled sources from excluded_relevant to chase, which identities to map, which papers to enrich, which talks to transcribe.

WAVE 2 (zoom): fire follow-ups via omniseek_gather, informed by WAVE 1 signals.

Structure: build an evidence graph from your findings (per server instructions section 6).

2. Investigation starting points

These are STARTING POINTS. Adapt, skip, extend based on what you find.

Person due-diligence

WAVE 1: omniseek_search + omniseek_resolve_identity. WAVE 2: omniseek_coauthors + omniseek_paper_enrich + walled chase + omniseek_transcribe. Key judgment: coauthor network reveals structural position; walled sources reveal candid views.

Lab / research group evaluation

WAVE 1: omniseek_search + omniseek_institution_cohort. WAVE 2: omniseek_field_skeleton + omniseek_coauthors(top PI) + omniseek_paper_enrich + walled chase. Key judgment: cohort shows who publishes there; citation neighborhood shows where they sit.

Field / topic mapping

WAVE 1: omniseek_search + omniseek_field_skeleton. WAVE 2: omniseek_paper_recommend + omniseek_paper_enrich + omniseek_transcribe. Key judgment: identify consensus core (high in_degree), frontier (recent, citing core), controversy.

Product / tool / company assessment

WAVE 1: omniseek_search(query="<product> review") + omniseek_search(query="<product> alternative"). WAVE 2: walled community sources + omniseek_read for official + critical pages. Key judgment: corroboration + also_in, conflicts, source_diversity (vendor-only = one-sided).

3. Saturation chase (walled-source depth pursuit)

After any broad search, read _meta.excluded_relevant. Each entry has overlap (query-token match count) and a sources=[...] re-run hint. Chase when: domain matches + overlap >= 2 + budget allows. Top 2-3, not all.

Fire-then-collect (per server instructions section 7) for parallel walled retrieval. If a xiaohongshu note URL appears, omniseek_read(url) retrieves full note + comment thread.

4. Budget discipline

• Surface (fast): 1 gather with 2-3 tools
• Standard: 2 gathers (sweep + zoom)
• Deep: 2-3 gathers + transcription
• Max 3 omniseek_search per gather (each fans out to many sources)
• Failed/empty/timed-out calls still count; do NOT infinite-retry

5. Triangulation and disconfirmation

• Count INDEPENDENT corroboration: corroboration counts source NAMES, not backends.

Cross-reference also_in to dedup mirrors. omniseek_graph(view="voices", args={"doc_ids": [...]}) is the tool-layer realization: hand it the doc ids from a search and it collapses them to distinct upstream voices via same_as + authored, so "5 sources agree" becomes "3 independent voices agree" (or fewer). The unresolved bucket is the honesty line: docs it could not resolve are returned separately, NEVER counted as voices. The other graph views (find -> stats -> neighborhood -> between -> since -> similar) walk the relation memory: since projects what accreted around a node after a date (stored edges only, tier + method shown, no collapsing), and similar lists the vector-nearest doc candidates for an anchor doc by rank (proposals only, method align:embed). When an exploratory view or similar surfaces a same_as candidate you verify, record it with omniseek_ruling(action=create) so the working policy collapses it thereafter. Relations judged FROM content (a filing says A acquired B) persist via omniseek_statement (directed, typed, attributed with note + source doc; identity types stay with omniseek_ruling).

• Flag conflicts, don't average: when _meta.conflicts appears, present the divergence,

don't pick a winner.

• Stamp freshness: carry each key fact's freshness_days.
• Reverse the query: before acting on a claim, search the INVERSE.
• Reference-class base rate: for "join this lab / take this path", estimate the denominator

with omniseek_institution_cohort before trusting anecdote.

6. Evidence graph construction

Build an evidence graph (structure in server instructions section 6):

• Extract claims conservatively (distinct, testable assertions with scope).
• Require sourced_from edges for every Claim.
• Use Phase A signals to inform edge weights and Gap identification.
• Check graph quality: orphan Documents, unsupported Claims, open contradicts, critical Gaps.

7. Close with a gap-ledger

End every investigation with an explicit accounting of what you did NOT get: sources used, queries sent, and the gap-ledger (zero-hit facets, unchased excluded_relevant, timed-out sources). These are VISIBLE gaps, never silently backfilled.