What the documents look like
An MCN agency manages long-term creator relationships through signed creator contracts, training records, violation handling forms and account schedules. These documents are updated daily and weekly, and they come from the MCN back office, platform violation notices, creator account data and the contract ledger. Product material and selection questions must keep the document subject, period and field source intact. FastGPT requires at least one language model and one index model to operate. From V4.17.0, AI Proxy is a required dependency; both the App and any Admin service that reuses App/Service configuration must set AIPROXYAPIENDPOINT and a non-empty AIPROXYAPITOKEN, and a missing item causes startup validation to fail.
Constraints at the model connections step
Creator contracts, training records, violation handling forms and account schedules are separate document types with inconsistent field naming and validity ranges. A question-answering result can only fall inside the retrieved source text, and retrieval does not perform authorization checks. A similarity score is a ranking basis and does not act as a subject, entity or permission filter. Libraries built on different vector models cannot be associated together; identifiers must be checked first and non-matching libraries excluded. Image retrieval requires a corresponding image input and model capability, and reranking requires a rerank model enabled by configuration. The optional paths shown in the overall document flow diagram do not mean every request executes all of them. Model call logs are retained for 1 hour, configurable through an environment variable, and historical requests cannot be reviewed after that window.
Hypothetical exercise sample: an MCN agency contract ledger contains a Creator Signing Contract whose party A is an agency, party B is creator A, signing date is 1 March 2026, contract term runs from 1 March 2026 to 28 February 2027, and the status column records "pending review". The same ledger contains an Account Schedule recording that creator A takes part in a brand special live session on 10 April 2026. The sentence to be judged is: "Creator A's signing contract has taken effect, so the April 2026 special live session can be arranged." Checking requires the contract review record and evidence of effectiveness separately. "Pending review" only describes the review stage; the material does not state the effectiveness conditions or whether they were fulfilled, so the claim "has taken effect" lacks support and the record remains to be verified. Signing, reaching cooperation and contract effectiveness are three distinct facts, and pending review does not show that operating permission is absent.
Configuration and implementation checks
The table below lists configuration check items for this page; trial values must be compared against actual deployment conditions.
| Configuration item | Hypothetical trial starting value | Reason and comparison method |
|---|---|---|
| AIPROXYAPIENDPOINT | http://aiproxy:3000 | Enter the AI Proxy service root address without appending /v1; compare against the internal service address in the deployment template |
| AIPROXYAPITOKEN | A non-empty string | Check the source of the administrator token and the server configuration; application API key usage is checked separately |
| AGENT_ENGINE | fastAgent | Options are fastAgent or piAgent; compare against the agent engine the team actually uses |
| HNSWEFSEARCH | 100 | Applies only to PG, OceanBase and openGauss; compare against the vector store type to confirm the applicable range |
| VECTORMAXPROCESS | 10 | Maximum concurrency of the vector training queue; compare against the parse queue concurrency to confirm each is independent |
| LLMREQUESTTRACKINGRETENTIONHOURS | 6 | LLM request tracking retention period; compare against the 1-hour retention of model call logs |
A model ID is the value of the model field in the request body and is globally unique; the display name can be customized. Model mapping requires separate checking of the upstream target.
Three common errors
- Treating a retrieval relevance score as a permission filter. A similarity score only affects ordering and does not exclude unrelated business records or unauthorized records. Retrieval is not authorization; business verification, approval, database writes and external synchronization each need their own downstream implementation. The check compares retrieved results against the access permission list and confirms the two do not substitute for each other.
- Reading pending review directly as effective. The sample holds only a review-stage record and lacks effectiveness conditions and fulfillment evidence, so contract effectiveness can be neither confirmed nor denied from it. The check compares the effectiveness claim in the drafted sentence against its supporting evidence; where absent, the item remains to be verified.
- Merging libraries built on different vector models. Cross-model association is not possible; identifiers must be checked first and non-matching libraries excluded, and no cross-model retrieval is run to compare missed recall. The check inspects the vector model identifier of each library and records the actual observation rather than inferring a merged runtime result.
How to verify the setup
- Input the contract source text, the account schedule and the drafted sentence, and check the review stage, effectiveness evidence and live session basis separately. The source text holds only pending review and a schedule entry, so it cannot support "the contract has taken effect"; that claim remains to be verified. The two stages are not treated as mutually exclusive.
- For retrieval scope, the input is the query "creator A April 2026 special live session", and the comparison object is the document identifiers in the returned results against the list of accessible documents. The expected observation is that returned results fall within the retrieval scope to which access is held; if an empty result is returned, use the explicit paper input
[]to check the downstream design and record the gap, without claiming that a fictional rule makes retrieval always return[]. - For model connection, the input is the configured values of AIPROXYAPIENDPOINT and AIPROXYAPITOKEN, and the comparison object is the actual server configuration. Check separately whether the address points to the target service and whether the administrator token matches the server; both being non-empty satisfies only these two required conditions, and overall startup and invocation still require actual verification. When a value is missing or empty, record the startup validation gap as documented. Where no execution has taken place, keep this as an expectation and do not claim it was tested.