What the documents look like
This page covers retail plastic bag and single-use item reporting, specifically the product information and selection question-answering task at the model connection and configuration step. The materials involved are plastic bag procurement ledgers, sales vouchers, semi-annual reports, and self-discipline commitments. These documents carry fields such as entity name, product specification, procurement or sales quantity, corresponding period, and voucher status. A question-answering setup must be able to cite these materials while treating any status not stated in the source as unconfirmed.
The reporting context differs from membership marketing or checkout promotion. The work revolves around disposable plastic product reporting and environmental supervision, with data drawn from the Ministry of Commerce reporting system, POS plastic bag sales, procurement systems, and store ledgers. The consequence of error is a deadline for correction, a fine between 10,000 and 100,000 yuan, or a regulatory interview. These facts shape what the question-answering feature must handle: it reports what the documents say and does not convert an unstated reporting or approval status into a confirmed one.
Constraints at the model connections step
Model access depends on AI Proxy. Both the App and Admin services must set AIPROXY_API_ENDPOINT and a non-empty AIPROXY_API_TOKEN; missing either causes startup validation to fail. Legacy OpenAI environment variables or a single-model custom request address cannot substitute for these two settings. AIPROXY_API_ENDPOINT takes the AI Proxy service root address without appending /v1. AIPROXY_API_TOKEN must match the server-side configuration and cannot be empty.
The system requires at least one language model and one index model to operate. Model types include language model, index model, rerank model, speech synthesis, and speech recognition. A model ID is the value of the model field in the request body and is globally unique. A model name is for display and can be customized.
Model status probing applies only to currently enabled system models. After a first probe failure, the system retries automatically up to 3 times, for at most 4 attempts in one round; only when all 4 attempts fail is the model judged abnormal. This mechanism observes model availability and does not change the authorization state of any material.
Retrieval and relevance locate text; they do not perform permission or entity isolation. The same-vector-model requirement for a retrieval library is limited to that library and does not extend to external systems. Libraries built on different vector models cannot be jointly associated; identifiers must be checked first and non-matching items excluded.
Configuration and implementation checks
The table below addresses the primary task of this page: model connection and configuration for product information and selection question-answering. Values marked as hypothetical starting points are test assumptions, not official defaults.
| Configuration item | Hypothetical test starting point | Notes and comparison method |
|---|---|---|
AIPROXY_API_ENDPOINT | Assume http://aiproxy:3000 | Chosen because the documented example uses this form without /v1; compare against the requirement that no /v1 suffix appears |
AIPROXY_API_TOKEN | Assume a non-empty string | Compare against the AI Proxy server-side configuration; confirm it matches and is not empty |
| Language model | Assume one language model enabled | Compare against the requirement of at least one language model |
| Index model | Assume one index model enabled | Compare against the requirement of at least one index model |
| Model status probe interval | Assume 30 minutes | Documentation supports an integer between 5 and 60 minutes; 30 is a middle value, compared against the allowed range |
| Alert Webhook address | Assume one HTTP address | Compare against the documented HTTP/HTTPS support; confirm the test button can be clicked |
Three common errors
- Filling
AIPROXY_API_ENDPOINTwith a complete model request address. Documentation requires the root address without/v1. The check takes the configured value as input, compares whether it contains/v1or/chat/completions, and treats either as non-conforming. - Using an application API key in place of
AIPROXY_API_TOKEN. Documentation requires an admin token consistent with the server side. The check takes the token as input, compares it against the AI Proxy server-side configuration, and treats a mismatch or empty value as non-conforming. - Treating a retrieval result as an authorization credential. Retrieval returns text and does not determine whether a material has been approved or reported. The check takes a question asking whether a given ledger has been reported, compares whether the answer claims it has been reported, and when the material does not state a reporting status, treats a claim of reporting as non-conforming.
How to verify the setup
Hypothetical paper exercise. Two source sentences from the material: A store's plastic bag procurement ledger states that in the first half of 2026, 12,000 plastic bags of specification 30cm×50cm were procured, the supplier is a packaging factory, and the ledger status is registered. The store's semi-annual report states that for the same period, sales vouchers total 11,500 units, and the report status is pending submission. Proposed determination sentence: The store has submitted the semi-annual plastic bag report for the first half of 2026.
- Input the source sentences above and the proposed determination sentence. Compare the reporting status in the proposed sentence against the report status field in the material. The material states pending submission; the proposed sentence states submitted. The two values are mutually exclusive on the same entity, same period, and same field, so the determination is a contradiction.
- Input the source sentences above and a question asking about the store's procurement quantity. Compare the quantity in the answer against the 12,000 units stated in the ledger. An answer giving 12,000 units and citing the ledger is supported; an answer giving another quantity or not citing a source is unsupported.
- Input a question asking about the store's sales voucher quantity. Compare whether the answer cites the sales voucher material. The material states 11,500 units; an answer citing that number and the voucher is supported; an answer not citing the voucher or giving an unstated number is pending verification.
All three items are paper checks and were not actually executed. Fields not stated in the material remain unknown; a pending submission status does not indicate submission, and a registered ledger does not indicate approval.