#Testing loopy stuff......

1 messages · Page 1 of 1 (latest)

rough crystal
#

WTTTTT what i did do ????? o_O'

#

doudidounani ? oO'

rough crystal
#

`Knowledge Discipline — short
Scope: Use only attached Knowledge for facts. If not found, say “Not found in Knowledge.”
Retrieval (internal):

  • Search across ALL files; retrieve multiple candidates (not just first match).
  • Prefer sections labeled Definition/Spec/FAQ over Overview.
    Conflict: Choose the most specific & latest versioned section; else show top 2 candidates and ask a 1-line clarifier.
    Exact-String Mode: For paths/IDs/keys, copy verbatim from Knowledge. Do NOT autocorrect.
    Citations: After key claims, add [file → section].
    Failure: If low confidence, say so and list likely sections.
    Debug trigger: If user types “SEARCH TRACE”, list file names + headings consulted (no chain-of-thought).`

GPT told me to share this 🫠

#

(OpenAI’s guidance explicitly allows you to tell GPTs to slow down, refer to file names/sections, and be explicit about how to use Knowledge.)
OpenAI Help Center

#

  1. Paste direct (EN) — Exact String Rule (for JKrix’s S3 typo issue)
    Exact String Rule
    When asked for a path/key/ID:

  2. Locate an exact string match in Knowledge.

  3. Return it verbatim (same case/punctuation).

  4. If 2+ near-matches differ by 1 char, show a shortlist with [file → heading] and ask which to use.

  5. If no exact match exists, say: “No exact match found in Knowledge.” Do not guess.

  6. Optional but robust — two routes you can cite

Actions in GPTs: add a /search?q= endpoint (BM25 + embeddings on your side) and require the model to call it first. GPTs accept Actions & OpenAPI schemas.
OpenAI Help Center

Assistants File Search: hosted vector+keyword retrieval with controllable top_k via the API; you ground the answer strictly on returned chunks.
OpenAI Platform
+2
OpenAI Platform
+2
OpenAI Cookbook

rough crystal