What the documents look like
Price and display auditing for chain stores happens during ordinary trading periods. The work centres on whether standards continue to be followed week by week and month by month, not on onboarding or launch coaching. The materials that reach document parsing and chunking are store visit photographs, price tag lists, display standards, and rectification notices. Two structural properties matter here. Price tag lists and head office promotion calendars are comparatively structured: tables or line items that can be split by field. Store visit photographs and display standards are mixed image-and-text materials, so whether an image description index or an image vector index is created at ingestion determines whether image retrieval can later reach them. The supplied brief gives weekly and monthly update frequencies. This page does not derive batch synchronisation, full synchronisation, real-time behaviour, or transaction-triggered mechanisms from those frequencies.
Constraints at the document chunking step
In FastGPT a knowledge base consists of library, collection, and data. The smallest search unit is the library; collections only classify and manage data and do not affect search results. One data item can map to multiple vectors, and when vectors resolve to the same original data item they are merged with the highest score retained. Whether an image can be retrieved depends on whether a corresponding index was built at ingestion, and on whether a vision model or an image-capable vector model is configured. Hybrid retrieval fuses multiple result paths through RRF. Reranking requires a model and must be enabled according to settings. Image retrieval requires image input and model capability. These paths are listed in the overall retrieval flow and do not all execute on every request. Minimum relevance is a filter threshold that takes effect only in semantic retrieval or when result reranking is used. It does not perform permission or entity isolation, and retrieval itself does not constitute authorisation.
Hypothetical paper exercise sample: a chain brand operations department holds a store visit record whose source text reads, "In week 2 of March 2025, the price tag list for store A-12 in district A showed 'pure milk 1L' at 9.9 yuan, and the head office promotion calendar for the same day recorded a promotion price of 8.9 yuan for that item, with the promotion period running from 10 March to 16 March." A draft sentence awaits review: "This store's price tag matches the head office promotion calendar, and price execution is correct." In this sample the subject is store A-12 in district A, the object is pure milk 1L, the state is that a tag price and a calendar promotion price coexist, and the effective scope is 10 March to 16 March. Comparing the draft sentence with the source text, the tag price of 9.9 yuan and the calendar promotion price of 8.9 yuan are different values on the same field within the same period, so the draft sentence's claim of a match is not supported. This sample is a paper exercise and was not submitted to retrieval or approval.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | Reason and comparison method |
|---|---|---|
| Search mode | Hybrid retrieval | Price tag identifiers and promotion calendar entries contain exact numbers, while natural language questions also need coverage; compare recall against semantic retrieval alone |
| Reference limit | 2000 tokens as the hypothetical starting value for this exercise | Store visit records and price tag lists are long per item; compare against a 1000 token starting value to see whether condition statements are truncated |
| Minimum relevance | 0.3 as the hypothetical starting value for this exercise | This value takes effect only in semantic retrieval or result reranking; compare against a 0.5 starting value to see whether unrelated fragments decrease |
| Question optimisation | Enabled | Follow-up questions about tag differences leave referents unclear; compare against the disabled state to see whether the search terms are completed |
| Image index | Build an image description index at ingestion | Store visit photographs need to participate in retrieval; compare against the state where no index is built to see whether an image query can hit |
| Chunking method | Split by price tag list line items and promotion calendar fields | Avoid separating a condition from the content it constrains; compare against whole-document chunking to see whether conditions stay complete |
Three common errors
- Treating a successful upload as an updated retrieval index. Input: a newly uploaded rectification notice. Compare the upload completion state against the actual training and index state, and observe whether retrieval can hit the notice content. A successful upload and an effective index are different facts.
- Treating minimum relevance as a permission filter. Input: a low-relevance price tag record belonging to another district. Compare the filter threshold against the access scope, and observe whether the record is excluded because of its low score. Relevance does not perform entity or permission isolation.
- Treating a rectification notice pending review as completed rectification. Input: a rectification request whose status is pending review. Compare the review status against the execution status, and observe whether a completion record exists. Pending review does not show that operating permission is absent, and it does not equal completed rectification.
How to verify the setup
- Against the hypothetical sample source text, enter "Are the price tag and calendar for pure milk 1L at store A-12 in district A consistent in week 2 of March". Compare whether the retrieved content contains both the tag price record and the calendar promotion price record, and observe whether both records share the same subject, the same period, and the same field. If only one record is returned, the claim is not supported and the gap must be recorded.
- Against the draft sentence "This store's price tag matches the head office promotion calendar, and price execution is correct", enter that sentence together with the source text. Compare the draft wording against the source values, and observe whether mutually exclusive values appear on the same field. If the tag price and the calendar promotion price coexist and differ, the draft sentence is not supported.
- Against the image retrieval path, enter a hypothetical store visit price tag photograph. Compare whether an image description index was built at ingestion and what the current model capability is, and observe whether the image query can hit. If no index was built or the model does not support image input, record this as not supported and do not infer that retrieval must return empty. An empty result can be checked against downstream design using the explicit paper input [], without claiming that a fictional rule retrieval necessarily returns [].