What the documents look like
Closure and liquidation work runs in the opposite direction from site selection and store opening. The material records an exit and the recovery of assets. Its main sources are POS inventory, the fixed-asset system, lease contracts and the finance ledger. The documents handled day to day are closure approvals, inventory clearance sheets, lease termination agreements and asset count sheets. Updates are irregular and occur per store.
In a cross-border setting, the lease termination agreement for one store may be signed in more than one language, the units of measure on inventory clearance sheets may differ, and the numbering scheme on asset count sheets may not match. Retrieval has to check the parties, the object and the valid scope against the source wording. A similarity score does not separate one store from another, and it does not separate records in different authorization states.
Constraints at the model connections step
FastGPT reaches model providers through AI Proxy. From V4.17.0, AI Proxy is a required dependency. Both the App and an Admin service that reuses App/Service configuration must set AIPROXYAPIENDPOINT and a non-empty AIPROXYAPITOKEN. A missing item causes startup validation to fail. Older OpenAI environment variables or a single-model custom request address do not replace these two settings. The system needs at least one language model and one index model to work normally.
Language model fields include maxContext, maxResponse, quoteMaxToken, maxTemperature, vision, toolChoice and functionCall. Index model fields include defaultToken and maxToken. Model status probing covers only currently enabled system models. After a first failure the system retries at most 3 times, so one round makes at most 4 attempts, and only 4 failed attempts mark the model abnormal. The probe interval accepts an integer between 5 and 60 minutes.
Hypothetical paper exercise sample: the lease termination agreement for one store of a chain brand reads, "The lessor and the lessee signed this agreement on 10 March 2026; the lease relationship terminates on 30 April 2026; refund of the deposit is conditional on completion of the asset count reconciliation." The asset count sheet in the same batch states, "Count date 15 April 2026; reconciliation not completed." The proposed determination sentence is, "The store lease has been terminated and the deposit can be refunded." The signing date and the termination date are different fields of the same agreement. The count sheet status is reconciliation not completed. The deposit refund has not entered execution.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | Reason and comparison object |
|---|---|---|
| AIPROXYAPIENDPOINT | The AI Proxy service root address | Compare with the documented format; confirm no /v1 is appended |
| AIPROXYAPITOKEN | A non-empty administrator token matching the server | Compare with the documented non-empty requirement; confirm it is not an application API key |
| Language model maxContext | 125000 | Compare with the combined length of the lease termination agreement and count sheet to be retrieved |
| Index model maxToken | 3000 | Compare with the chunk length of a single inventory clearance sheet |
| Probe interval | 30 minutes | Compare with the documented integer range of 5 to 60 minutes |
| ReRank model | Enabled as needed | Compare with the documented condition that a model must exist and be enabled by configuration |
Three common errors
- Treating a similarity score as an authorization filter. A retrieved lease termination agreement may score close to a record from another store while the parties differ. Check: input the sample wording above together with a second store identifier, compare the store identifier on the returned records, and treat any cross-store record as not conforming.
- Treating the signing date as proof that termination has taken effect. In the sample the signing date is 10 March 2026 and the termination date is 30 April 2026; these are different fields of the same agreement. Check: compare the proposed determination sentence with the source wording and record "has been terminated" as unsupported.
- Treating an incomplete count reconciliation as a refundable deposit. Check: input the count sheet wording and the proposed determination sentence, compare the status field, and treat a "can be refunded" conclusion as not conforming.
How to verify the setup
- Verification one: input the sample lease termination agreement wording and the proposed determination sentence, "The store lease has been terminated and the deposit can be refunded." Compare the signing date, the termination date and the count sheet status. Only when the three fields share the same party, the same period and clearly mutually exclusive values in the same field is a contradiction recorded; otherwise record the result as unsupported. The expected observation is that the deposit refund condition remains unmet because reconciliation is not completed.
- Verification two: use the explicit paper input
[]to check the downstream design. Observe whether the empty-result handling path behaves as designed. Do not claim that a search under an invented rule necessarily returns an empty array. - Verification three: check the index model identifier and exclude records whose vector model does not match the retrieval library. Record only the returns actually observed. Do not run a cross-model retrieval to compare missed recall.
After configuration, the model status page refreshes automatically every 1 minute while automatic probing is on. The alert Webhook pushes one notification per model when the status changes at an edge, with event types modelstatuserror and modelstatusrecovered.