How Find Anything works

Overview

Find Anything sits between your AI agent and your knowledge sources. It enforces permissions, retrieves candidate evidence, ranks it, detects conflicts, and returns a structured result your agent can reason about — including an honest account of what it could not read.

Retrieval API

A single request returns the answer with its citations, coverage, conflicts, confidence, and a receipt_id.

// Request
{
  "query": "current enterprise price",
  "scope": "agent:acme-sales",   // permissions enforced server-side
  "connector": "google_drive"
}

{
  "answer": "Enterprise: $120,000/yr + $25,000 impl.",
  "confidence": "medium",
  "citations": [
    { "source": "Pricing Final.pdf", "page": 2, "authoritative": true },
    { "source": "Pricing Draft.pdf", "page": 1, "superseded": true }
  ],
  "conflicts": [
    { "field": "annual_price", "values": ["$120,000", "$95,000"] }
  ],
  "coverage": { "known": 8, "searched": 7,
                "excluded": [{ "reason": "out_of_scope" }] },
  "receipt_id": "rcpt_9f2a…"
}

Coverage accounting

Coverage is the difference between what is known and what was read. known counts sources whose existence is already recorded — inventoried during a previously authorized sync; searched is how many were actually retrieved; excluded lists what was skipped and why (authorization lost, out of scope, unavailable, unsupported format). Restricted content is never fetched, extracted, indexed, quoted, or returned, and Find Anything cannot discover files the connected account can't enumerate.

• Honest denominator — results report 7 of 8, never "complete".
• Distinct outcomes — searched, excluded (out of scope), unavailable, unsupported, and failed are separate states.

Conflicts & authority

When two in-scope sources disagree on a field, the disagreement is reported — both values, both citations. Ranking never deletes the loser. Authority is explained from observable signals: recency, an approved status, and explicit "final/authoritative" markers. Find Anything explains why one source ranked higher; it does not claim the higher-ranked value is objectively true.

Confidence

Confidence summarizes how much to trust a result. It drops when conflicts are unresolved, when coverage is incomplete, or when the top source lacks authority signals. It is a signal for your agent's control flow — e.g. escalate to a human when confidence is low.

Replay, receipts & diffs

• Receipt — a stored record of what a retrieval ran against: document-version and index snapshots, the permission context, extraction and ranking configuration versions, and the query parameters.
• Replay — re-run a stored receipt. Against unchanged inputs the result is deterministic. Replay does not promise that future runs match once documents, permissions, or software versions have changed.
• Diff — when inputs have changed, a re-run produces a comparison of the two runs (e.g. the day Pricing Final superseded Pricing Draft), not a claim of the same result.

Boundaries & connectors

The Google Drive integration is built and tested, working in private beta; authenticated access is pending Google restricted-scope verification. Everything else is labeled Planned and is not available yet. Find Anything only ever reads what the agent is already permitted to read.

← Back to overview