What the documents look like
After-sales repair and reship work aims to restore a product to use. Cases close through repair completion, reship dispatch, and customer confirmation; a direct refund follows a separate handling path. The daily materials come from the after-sales system, the order system, repair outlets, and the spare-parts warehouse WMS. Case progression is per order, while parts and outlet status update daily. The documents entering contract and transaction material comparison include repair requests, reship orders, inspection reports, and spare-part outbound orders. These four material types differ in structure: repair requests and reship orders are form-oriented, inspection reports are conclusion paragraphs, and outbound orders are lines of identifiers and quantities. Document parsing and chunking must let the same material, after segmentation, still point back to the document identity and version of the source. Otherwise later comparison loses its object.
Constraints at the document chunking step
This section gives one hypothetical drill sample to explain how parsing and chunking are checked. It points to no real customer and no real document.
The source text supplied for checking in the hypothetical drill sample is three sentences. Sentence one: "Reship order number BJD-2024-0317, linked order DD-88231, application date 17 March 2024, reship object is one power adapter." Sentence two: "Inspection report number JCBG-2024-0317, inspection conclusion is abnormal adapter output, replacement recommended." Sentence three: "Spare-part outbound order number CKD-2024-0318, outbound quantity one, outbound date 18 March 2024, receiving outlet is Chengdong repair point."
One draft sentence to be judged is: "The order has completed reship, and the part has been signed for by Chengdong repair point." Judging this draft sentence requires evidence of the same subject, the same period, and the same field. The materials contain a reship order, an inspection report, and an outbound order. The outbound order records only outbound quantity and receiving outlet. It has no signature field and no reship completion status field. The draft sentence is therefore unsupported; the signature fact remains to be verified, and reship completion cannot be concluded from it.
Chunking constraints follow from this structure. Reship order number, order number, and application date belong to the same form subject, and a condition and the content it constrains cannot be split apart. The inspection report conclusion sentence needs to stay in the same data group as the report number and date. In FastGPT, a knowledge base consists of library, collection, and data. The smallest search unit is the library, and collections are used for classification management. One data group can map to multiple vectors, and the indexes field records an array of vector IDs. When the same source data is recalled through multiple paths, it is merged and the vector score takes the highest value. These mechanisms show that chunk quality directly affects whether recall brings out the complete subject.
Configuration and implementation checks
The table below lists documented settings or implementation checks around this page's sample. The values are starting points for this hypothetical trial, used to explain the comparison method, and are not official default values.
| Configuration item | Hypothetical trial starting point | Reason and comparison object |
|---|---|---|
| Search mode | Hybrid search | The sample contains identifiers and natural-language conclusions, so both input types need coverage |
| Citation limit | Hypothetical starting point 3000 tokens | The sample's three sentences are short, and the result is compared with the full chunk length when not truncated |
| Minimum relevance | Hypothetical starting point 0.4 | Used to observe whether chunks containing identifiers remain after filtering, compared with a lowered value |
| Question optimization | Enabled | The sample question contains reference, and recall content is compared with the disabled state |
| Result rerank | Enabled | Used for secondary ordering when candidates are many, compared with ordering without rerank |
| Image retrieval | Not enabled | The sample has no image input, no vision model, and no image vector index |
The citation limit is controlled by tokens. Minimum relevance takes effect only under semantic search or when result rerank is used, and it is a filtering threshold. RRF is the fusion process for multiple result paths in hybrid search, and it is not configured as a standalone switch.
Three common errors
- Treating the receiving outlet on the outbound order as signature evidence. The input is the outbound order source text; the compared fields are "receiving outlet" and "signed for" in the draft sentence; the observable result is that the outbound order has no signature field and the draft sentence is unsupported.
- Treating a successful upload as retrieval already updated. The input is the upload action after a material change; the comparison object is the actual training and index state; the observable result is that while the index state is unconfirmed, retrieval may still return the old chunk.
- Using minimum relevance to carry permission or entity isolation. The input is the minimum relevance setting; the comparison object is the subject identifiers in recalled results; the observable result is that this threshold does not exclude unrelated subject records, and retrieval does not constitute authorization.
How to verify the setup
- Using sentence one of the sample as input, retrieve reship order number BJD-2024-0317, and compare whether the returned chunk contains both the order number and the application date. If the chunk is missing subject fields, the condition is not met.
- Using sentence two of the sample as input, retrieve inspection report number JCBG-2024-0317, and compare whether the returned chunk contains the report number and the inspection conclusion. If the conclusion and the number are split into different data groups, the condition is not met.
- Using the sample draft sentence as input, compare whether the returned content contains a signature field. If all returned materials have no signature field, record that fact as pending verification and do not judge reship as completed.
If the return is empty, a clearly written paper input empty array can be used to check how the downstream design handles an empty result. This does not imply that a fictional rule certainly returns an empty array. Retrieval observation needs to record failure branches, including no hit, hit on an unrelated chunk, and hit on an incomplete chunk. Libraries across different vector models cannot be associated together. During checking, verify the identifier first and exclude those that do not match.