AGENT SKILL

enterprise-artifact-search

Multi-hop evidence search + structured extraction over enterprise artifact datasets (docs/chats/meetings/PRs/URLs), with strict question grounding, anti-distractor filtering, evidence-based entity selection, ID validation, JSON-ready outputs, and token logging that avoids unsupported installs.

License 许可
Apache-2.0
最近更新
2026-11-14
Stars
★ 0
Forks
1

01 前言 Front Matter

name
enterprise-artifact-search
description
Multi-hop evidence search + structured extraction over enterprise artifact datasets (docs/chats/meetings/PRs/URLs), with strict question grounding, anti-distractor filtering, evidence-based entity selection, ID validation, JSON-ready outputs, and token logging that avoids unsupported installs.

03 SKILL.md 正文

Steps

  1. Ground on the question text (no heuristics):
    • Read /root/question.txt and treat each entry (e.g., q1, q2, …) as the only specification for what to retrieve.
    • For each question, explicitly extract: target product/topic, required entity types (IDs/names/URLs/etc.), and any constraints (“author”, “key reviewers”, “competitor demos”, “from Slack”, “latest”, etc.).
    • Do not infer selection criteria from artifact structure (e.g., do not treat “meeting participants” as “reviewers” unless the question says so).
  2. Build a candidate artifact set (wide recall, then filter):
    • Start with product-scoped containers when available (e.g., /root/DATA/products/<Product>.json).
    • Expand to global search only as needed (grep/ripgrep over /root/DATA for exact strings from the question: product name, report title, competitor name, URL fragments, doc IDs).
    • Follow references across artifacts (slack ↔ doc links ↔ meeting transcripts ↔ PRs) to reach primary sources.
  3. HARD grounding & anti-distractor gate (prevent cross-product leakage):
    • Accept a candidate artifact only if it passes 2+ independent grounding signals relevant to the question’s target scope, such as:
      • Located under the correct product container and referenced from product-specific channels/meetings.
      • Title/content explicitly mentions the target product/topic.
      • Appears in a clearly product-scoped channel/meeting series.
      • Link/path includes a product-specific identifier consistent with the target product.
    • Reject artifacts that primarily concern a different product/topic unless the question explicitly requests cross-product comparisons.
  4. Select the correct “version” when multiple candidates exist:
    • Prefer explicit latest/final markers; else use most recent date; else use the one most referenced in follow-up discussions tied to the same grounded scope.
  5. Extract entities with evidence-based rules (no role assumptions):
    • Authors: prefer explicit fields (author/authors/created_by/owner), then PR author, then the Slack poster only if it clearly introduces the exact artifact.
    • Reviewers/contributors: use explicit reviewer/approver fields first; else include only people who provide substantive feedback in doc comments, meeting transcript turns, or Slack thread replies that are clearly tied to the artifact.
    • URLs/items/IDs: extract directly from artifact fields or message text; de-duplicate while preserving order.
  6. Normalize + validate outputs:
    • Normalize to the requested type (e.g., employee IDs as eid_...).
    • Validate every returned employee ID exists in /root/DATA/metadata/employee.json (or the dataset’s employee directory).
    • Ensure the final answer for each question is always a list (even length-1).
  7. Write results in the required JSON shape:
    • Write /root/answer.json as:
      • top-level dict keyed by question key ("q1", "q2", …),
      • each value is {"answer": [...], "tokens": "..."}.
    • Keep tokens as a string, matching the task’s example schema.
  8. Token logging (avoid failing installs; prefer available tooling; be explicit if approximate):
    • Do not attempt to pip install tokenizers in restricted environments.
    • First try to compute token counts using already-available libraries (e.g., python -c "import tiktoken"). If available, use it to count tokens for the per-question serialized payload you produce.
    • If no tokenizer library is available, record a transparent string value such as "approx:<n>" based on a consistent local heuristic (e.g., character-count/4), rather than implying an exact “consumed tokens” measurement.
    • Never store numeric token values; always store the token field as a string.

Expected Result

For every question in /root/question.txt, /root/answer.json contains an entry like:

  • "answer" is a list (deduplicated, validated, grounded in artifacts),
  • "tokens" is a string computed via an available tokenizer when possible, otherwise transparently marked as approximate (e.g., "approx:123"), and entity selection is driven by explicit question requirements and direct artifact evidence (not assumptions like “participants == reviewers”).

04 同仓库更多 Skills