What the documents look like
Weighing and sorting happen inside the warehouse fulfilment stage, so the working records centre on weight and specification. A sorting task sheet carries the order number, product name, target weight range and sorter. A weighing label is produced by the electronic scale and carries the actual weight, print time and barcode. A short-weight compensation form records the difference weight, compensation method and approval status. A quality inspection record carries photographs and a verdict. The operating entity is the front warehouse; the object is a specific order item; status fields include pending sorting, weighed, reviewed and pending approval. One order can correspond to several weighing labels, and each label weight must be compared line by line against the target range on the task sheet. The procurement and supplier material stage adds supplier name, purchase batch number and arrival date, which relate to the product code on the sorting task sheet. Chunking must keep the full context of that code.
Constraints at the document chunking step
The weight range on a sorting task sheet is the basis for judging a short weight. If chunking separates the range values from the product name, the retrieved fragment cannot support the comparison. The barcode and actual weight on a weighing label belong to the same source record; the multi-vector mapping in FastGPT allows one data record to correspond to several vectors, and retrieval recalls that record when any one vector is hit, so chunking should keep the fields inside a label together. On a short-weight compensation form the approval status and the compensation amount sit in the same passage, and splitting them causes status and amount to be mismatched. The photograph description index on a quality inspection record depends on a vision model, while the image vector index depends on a vector model that supports images; these are different conditions, so chunking must distinguish whether a text index or an image index was built. A purchase batch number in supplier material suits full-text retrieval, while a weight value suits semantic retrieval, and hybrid retrieval fuses the two paths through RRF.
Hypothetical paper exercise sample: the source text to be checked consists of three sentences. "Sorting task sheet ST-2024-0318-017, product spinach 500g±20g, sorter Zhang San, status weighed. Weighing label WB-2024-0318-017-01, actual weight 468g, print time 2024-03-18 09:12. Short-weight compensation form BC-2024-0318-017, difference weight 32g, compensation method refund 2 yuan, approval status pending approval." The draft sentence to be judged is: "This order has completed short-weight compensation approval and executed the refund." The subject needed for the judgement is order ST-2024-0318-017, the object is compensation form BC-2024-0318-017, the status field is approval status, and the effective scope is 2024-03-18. The source text does not state approval passed or refund executed, and the draft sentence writes pending approval as completed, so it is unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting value | Reason and comparison method |
|---|---|---|
| Chunk granularity | One task sheet or one label per segment | Compare with fixed-character splitting; check whether the weight range and product name stay in the same segment |
| Index type | Text index plus full-text index on the barcode field | Compare with semantic retrieval only; run an exact query on the batch number and observe hits |
| Reference limit | Hypothetical trial starting value of 2000 tokens | Compare with 1000 tokens; observe whether the compensation form fields enter the reference completely |
| Minimum relevance | Hypothetical trial starting value of 0.3 | Compare with 0.5; observe whether the pending-approval fragment is filtered out |
| Search mode | Hybrid retrieval | Compare with pure semantic retrieval; query the identifier "BC-2024-0318-017" and observe recall |
| Image index | Built according to model capability | Compare with a text-only index; check whether the inspection photograph description participates in retrieval |
Three common errors
- Splitting the actual weight on a weighing label from the target range on the task sheet into different chunks. The input is the label text, the comparison object is the range field on the task sheet, and the observable result is whether a query for "468g" also recalls the range "500g±20g". If the range is not returned with the weight in the same segment, the short-weight judgement lacks its basis.
- Treating the "pending approval" status on a short-weight compensation form as identical to "refund executed". The input is the compensation form text, the comparison object is the approval status field, and the observable result is whether the text returned for that form number contains "pending approval". If the returned text contains "completed", it contradicts the source status field.
- Treating the image description index and the image vector index as one condition. The input is the quality inspection photograph, the comparison object is the index type built at ingestion time, and the observable result is whether a pure image query returns a hit. If no image index was built, an empty image-search result is a missing index and is unrelated to the search parameters.
How to verify the setup
- Verification one: input the task sheet text from the hypothetical sample, compare whether the retrieval result contains both the product name "spinach" and the weight range "500g±20g", and observe whether both are returned inside the same chunk. The condition is met when they appear in the same segment; if not, adjust the chunk granularity.
- Verification two: input the compensation form number "BC-2024-0318-017" from the hypothetical sample, compare the approval status field in the returned text against the source value "pending approval", and observe whether "completed" or "refunded" appears. The condition is met when the status field matches the source; if not, check whether chunking separated the status from the amount.
- Verification three: input the draft sentence "This order has completed short-weight compensation approval and executed the refund" from the hypothetical sample, compare the field values in the draft against the three source sentences, and observe whether the source states approval passed or refund executed. The condition is met when the draft sentence is judged unsupported, because the source states only pending approval. For an empty-result scenario, a clearly written paper input of an empty array can be used to check the downstream design, without claiming that retrieval necessarily returns empty. Libraries across different vector models cannot be associated together; first check the vector model identifier, exclude those that do not match, and record the actual observation.