What the documents look like
Fresh-food supplier claims pursue quality responsibility for goods on arrival. They do not handle ordinary in-store shrinkage. The evidence chain runs through quality exception forms, photographic evidence, deduction notices and claim reconciliation statements. A single delivery may carry Chinese inspection records alongside English or local-language documents issued by an overseas supplier, so whether a claim holds depends on whether the responsible party, batch, time window and liability fields can be aligned across languages.
Materials appear at a per-batch or monthly rhythm. Retrieval work concentrates on grouping multilingual material under one claim matter and on recording what state each item had reached at handover. The update frequency does not change anything on this page, so it is left aside.
Constraints at the model connections step
Model connections decide whether multilingual material can be parsed into searchable text. FastGPT connects to model providers through AI Proxy. Both the App and any Admin service that reuses App/Service configuration must set AIPROXY_API_ENDPOINT and a non-empty AIPROXY_API_TOKEN; missing either one causes startup validation to fail. Older OpenAI environment variables or a single-model custom request address do not replace these two settings. A system needs at least one language model and one index model to work normally. The index model builds indexes over text chunks for related-text retrieval.
Semantic recall and relevance scores carry no permission or entity-isolation role. A retrieval library requires association with the same vector model; this requirement does not extend to external systems such as business-registration sites.
Hypothetical paper exercise. Inspection record for one fresh-produce delivery, original wording: "12 March 2026, batch A-0312, sampling found a high proportion of wilted leafy vegetables, judged as a quality exception on arrival, responsible direction is the supplier." English deduction notice for the same batch, original wording: "Deduction Notice: Batch A-0312, issued 2026-03-15, deduction amount pending supplier confirmation." One draft sentence awaiting judgment: "The supplier has confirmed the deduction amount for this batch." That draft sentence concerns a party, an object and a state, and can only be judged after comparison with both source texts. The deduction notice states the amount is pending supplier confirmation. The draft sentence asserts confirmation. The two do not agree.
Configuration and implementation checks
The table below works through the exercise item by item. Each starting value is a hypothetical trial starting point for this page, used for comparison against a target value. None of these values is an official default.
| Configuration item | Hypothetical starting point | Comparison method and reason |
|---|---|---|
| AIPROXYAPIENDPOINT | Hypothetical AI Proxy service root address | Compare with the deployment document example; confirm no /v1 is appended |
| AIPROXYAPITOKEN | Hypothetical non-empty admin token | Compare with server-side configuration; confirm non-empty and consistent |
| Language model | Hypothetical one enabled language model | Check together with the index model against the documented minimum |
| Index model | Hypothetical one enabled index model | Check that text-chunk indexing is available for related-text retrieval |
| Model probing | Hypothetical automatic probing on, interval an integer between 5 and 60 minutes | Compare with the documented range; observe the status chart and alerts |
| Retrieval library vector model | Hypothetical single vector model per retrieval library | Check identifiers before association; do not associate across models |
Three common errors
- Treating semantic recall as a permission filter. Input: the draft sentence and the retrieval library. Compare relevance against authorization state. The observable result is that relevance does not change whether material has been authorized.
- Writing a pending state as completed. Input: the deduction notice stating "pending supplier confirmation" and the draft sentence stating the supplier has confirmed. Compare the two. The observable result is that the draft sentence is unsupported by the source wording.
- Associating across models. Input: two retrieval libraries using different vector models. Compare their identifiers and exclude the non-matching one. Record what is observed; do not infer missed recall or redundant results.
How to verify the setup
- Input the two source texts and the draft sentence from the exercise above. Compare party, batch and state fields. If the draft sentence's "has confirmed" has no corresponding wording in either source, the draft sentence is unsupported and the evidence gap should be recorded. The deduction notice gives a status of pending, which is not the same field value as confirmed.
- Input the actual values entered for
AIPROXY_API_ENDPOINTandAIPROXY_API_TOKEN. Compare them with the deployment document requirements. If both are non-empty and the root address has no/v1appended, the startup validation condition is met; if either is missing, it is not met. - Input a clearly labelled paper empty result
[]. Compare how the downstream design receives an empty result and observe whether it treats the input as empty. This check does not claim that any invented rule retrieval necessarily returns[].