Fresh-food supplier claims: model connections for product questions

What the documents look like

Fresh-food supplier claims pursue responsibility for quality at arrival. They do not handle ordinary in-store shrinkage. The working documents are quality exception notes, photo evidence, deduction notices and claim reconciliation statements. These materials move on two rhythms: per arriving batch, and per monthly reconciliation. For one batch, the inspection conclusion, the accepted quantity and the deduction amount need to stay consistent with each other.

Product information and selection questions therefore have a narrow job. A question asks which material supports a given claim record, which stage that record has reached, and who currently holds it. The answer has to stay inside what the quality exception note, the photos, the deduction notice and the reconciliation statement actually say. A deduction notice that has been issued is not a settled claim. A supplier that has not confirmed is not a supplier that has agreed. These distinctions matter because the consequence of getting them wrong is supplier dispute, absorbed bad-goods cost, or interrupted supply.

Constraints at the model connections step

Model access runs through AI Proxy. For self-hosted deployments, AIPROXY_API_ENDPOINT holds the service root address and does not carry /v1. AIPROXY_API_TOKEN holds a non-empty administrator token that must match the server configuration. The system needs at least one language model and one index model to work. A model ID is the value of the model field in the request body and is globally unique; the model name is for display only. Model mapping connects a FastGPT-side model ID to the upstream model actually called, for example mapping gpt-4o-test to gpt-4o. Channel priority runs from 1 to 100, and a larger value is more likely to be selected first.

At the retrieval layer, the hnsw.ef_search parameter applies only to PG, OceanBase and openGauss, and HNSW_MAX_SCAN_TUPLES applies only to PG. A relevance score is a ranking input. It does not act as a permission filter or an entity filter, and retrieval is not authorization. Libraries built on different vector models cannot be associated together. When checking, compare model identifiers first, exclude libraries that do not match, and record what was actually observed.

Model capability is checked against the input the application needs. Multimodal input and tool calling are verified separately. Application parameters are set in the application or node configuration; the same-named fields of a compatible chat interface cannot be assumed to override the orchestration. Retrieval and forms collect and return information. Business verification, approval, database writes and external synchronisation each need their own downstream implementation.

Hypothetical drill sample. A quality exception note for one arriving batch reads: "Batch A-0312 arrived 12 March 2026. Sampling inspection failed. The deduction notice has been issued. The supplier has not confirmed." The draft sentence under review reads: "This batch claim is closed." The materials state only that the deduction notice was issued and the supplier has not confirmed, so the draft sentence goes beyond what the materials support.

Configuration and implementation checks

The values below are hypothetical starting points for this exercise, used only to describe a comparison method. They are not official defaults.

Three common errors

1. Writing AIPROXY_API_ENDPOINT as a full request address that includes /v1. The input is the environment variable value; the comparison is against the documented root-address format; the observable result is whether /v1 or /chat/completions was appended.
2. Treating a relevance score as a permission filter. The input is the score list returned by one retrieval; the comparison is between the returned records and the records the current account may see; the observable result is whether out-of-scope records appear.
3. Treating an issued deduction notice as a closed claim. The input is the hypothetical sample text and the draft sentence; the comparison covers the subject, batch and status fields; the observable result is whether the draft sentence exceeds what the materials state.

How to verify the setup

1. Input the hypothetical sample text and the draft sentence "This batch claim is closed." Compare the status fields the materials state. The materials record only that the deduction notice was issued and the supplier has not confirmed, so the draft sentence is unsupported and should be recorded as pending verification.
2. Input the actual value of AIPROXY_API_ENDPOINT and compare it against the documented format requirement. A value that does not contain /v1 and that passes startup validation satisfies the condition; a value containing /v1 is recorded as a configuration mismatch.
3. Input the records returned by one retrieval and the records visible to the current account, then compare the two sets. If every returned record falls inside the visible range, the condition is satisfied; if a record outside the range appears, the permission check is recorded as failed. For an empty result, a paper input of [] can be used to check the downstream design; no retrieval behaviour is inferred from it.

Across all three checks, the evidence gap stays open. The sample does not establish supplier agreement, settlement, approval or any completed business action, and none of these checks should be read as having been executed.

Related reading

• self-host/config/env
• self-host/config/model/intro
• FastGPT model connections