What the documents look like
Cold-chain receiving records are organised around a delivery batch. Four document types appear in this category: cold-chain temperature records, arrival inspection sheets, rejection notes, and supplier delivery notes. Temperature records describe the transport stage. Inspection sheets describe the store-level count and condition check. Rejection notes describe how non-conforming goods are handled. Delivery notes describe what the supplier declared. These documents are produced per batch, and store receiving and quality inspection staff repeatedly refer back to the same batch.
Document structure directly affects parsing results. A single batch temperature record may contain several timestamped readings. An inspection sheet may contain several item lines. If a parsing step separates a condition statement from the items it constrains, the retrieved fragment lacks the basis for judgement. The chunking task keeps each segment traceable to its source identity and version so that product information and selection questions can locate a specific field of a specific batch.
Cold-chain receiving differs from fruit procurement, which concerns origin price enquiries. This category concerns transport temperature and arrival batch, and the main data sources are the temperature-controlled vehicle, WMS, the store receiving system, and quality inspection photos. Errors here can lead to spoiled ingredients, supplier disputes, and regulatory penalties. Data updates per batch and per day. The three most time-consuming steps are measuring temperature, sampling quality, and registering the rejection reason.
Constraints at the document chunking step
Chunking follows the structure of each document type. Temperature timestamps, inspection item lines, and rejection reason descriptions each form a relatively complete semantic unit. Keeping a constraint together with the content it constrains allows retrieval to return subject and scope together. Original data and retrieval indexes are distinct; one set of original data can correspond to multiple indexes, and adding indexes does not change the original data content.
Retrieval offers semantic search, full-text search, hybrid search, and result reranking. Batch numbers, item numbers, and proper nouns suit full-text search. Natural-language descriptions suit semantic search. Hybrid search merges multiple result paths through RRF, and reranking applies a second ordering to candidate text. Whether these paths execute depends on the input and the settings; the optional paths shown in the document flow diagram do not indicate that every request executes all of them.
A hypothetical paper exercise sample: a store receives a batch of chilled dairy products. The source material for comparison reads: "The transport temperature record for this batch shows 2 to 6 degrees Celsius, with arrival at 9:20 on 8 April. The inspection sheet states the item is pasteurised milk, 200 cartons, of which 12 cartons have damaged outer packaging. The rejection note draft states the rejection reason as packaging damage, and the handling status as awaiting quality inspector signature." A draft sentence to be assessed reads: "This batch has completed rejection and the supplier has been notified to replenish." The assessment compares the handling status field of the rejection note with the completion wording of the draft sentence.
Configuration and implementation checks
Configuration centres on parsing and retrieval checks for this sample. The values below are starting points for this hypothetical exercise only, used for comparison against observed retrieval results, and do not represent official defaults.
| Configuration item | Hypothetical starting point for this exercise | Check method |
|---|---|---|
| Chunking basis | Split by document structure: temperature records by timestamp, inspection sheets by item line | Inspect whether each chunk contains subject and constraint |
| Index and original data | One set of original data can correspond to multiple indexes | Check whether original data content and index targets match |
| Search mode | Full-text for batch numbers, semantic for natural-language descriptions | Enter a batch number and a description sentence separately, observe returned content |
| Reference limit | Measured in tokens; starting point set to cover the complete item lines of one inspection sheet | Compare with a result covering only part of the item lines |
| Minimum relevance | Starting point set to a mid value in the 0-1 range | Compare with recall scope after raising and lowering |
| Update check | After material changes, check actual training and index status | Compare upload records with retrieval results for synchronisation |
Three common errors
- Treating a successful upload as retrieval already updated. Upload and training/index status are separate stages. After a material change, the actual training and index status must be checked, and retrieval results observed to see whether they reflect the new content.
- Treating retrieval results as an authorisation basis. Retrieval returns content fragments; it does not indicate that a rejection has been approved or that a batch has completed handling. The awaiting-signature status of the rejection note draft and the completion wording of the draft sentence are different facts, and the handling status field must be compared.
- Treating RRF merging in hybrid search as a standalone switch. RRF is a merging process for multiple result paths. Whether a merged result appears depends on which recall paths actually execute, and it cannot be assumed that every request passes through all paths.
How to verify the setup
- Verification one: Enter the handling status "awaiting quality inspector signature" from the rejection note draft in the hypothetical sample, and compare it with the draft sentence "This batch has completed rejection and the supplier has been notified to replenish." If neither the retrieval result nor the draft sentence records a completed signature or a notification record, the completion wording of the draft sentence is unsupported and should be recorded as pending verification.
- Verification two: Enter the batch number and the keyword "packaging damage", using full-text search and semantic search separately, and compare whether the returned content contains the item line of the inspection sheet and the reason field of the rejection note. If one mode does not return the corresponding field, record that mode's observed result without inferring that the other mode necessarily returns it.
- Verification three: After a material change, enter the field values from before and after the change for retrieval, and compare the returned content with the change records. If the returned content still shows the pre-change field value, record the difference between the upload record and the retrieval result, and do not treat a successful upload as an index already updated.
This page addresses only store cold-chain receiving, product information and selection questions, and document parsing and chunking. During operational handover, which materials, operations, and requests are in what authorisation state cannot be treated as authorisation execution through retrieval or forms. Semantic recall and relevance do not carry out permission or entity isolation. A retrieval library using the same vector model requirement cannot be extended to require the same model for business registration sites or other external systems. Any system error or automatic business action not stated in the source material must not be asserted. RRF is a hybrid result merging process and is not written as an enabling switch.