What the documents look like
Cold-chain receiving at a store is organized around a delivered batch. Each batch arrives with a cold-chain vehicle temperature record, an arrival inspection sheet, a rejection note, and a supplier delivery note. The parties are the store receiving side and the supplier; the object is the specific batch of goods; the states include accepted into stock, rejected, and pending re-inspection. Temperature records and inspection sheets carry time points, temperature values, and batch numbers. Rejection notes carry a rejection reason and a quantity. Delivery notes carry the supplier name and delivered quantity. Materials update per batch, and some data is summarized daily.
Document parsing and chunking must preserve the constraints between these fields. A temperature condition and an inspection conclusion cannot be separated, and a rejection reason cannot be placed in a different data block from its batch. If a condition clause and a conclusion clause are split during ingestion, a search can return the conclusion while losing the temperature premise, and the check loses its basis.
Constraints at the document chunking step
Chunking follows the structure of the material. The temperature record, sampling conclusion, and rejection reason of one inspection sheet should stay in the same data block or have a traceable association. Data and index are distinct: the MongoDB dataset.datas table stores the original data, and the indexes field records the corresponding vector IDs, so one group of original data can correspond to multiple vectors. During retrieval, vectors are recalled first, then the vector IDs are used to find the original data in MongoDB; entries that map to the same original data are merged and the highest score is taken.
Hypothetical drill sample: a store receives one cold-chain delivery. The source material reads: "On 10 March 2025, the cold-chain vehicle temperature record shows 2 to 6 degrees Celsius throughout transport. The arrival inspection sheet states batch A-0310 totals 40 cartons, 6 cartons were sampled, and sensory inspection passed. The supplier delivery note states the batch was delivered by Party B, quantity 40 cartons. No rejection note was issued that day." The draft sentence to be assessed is: "Batch A-0310 has completed inspection and has been fully received into stock." This sentence merges the sampling pass on the arrival inspection sheet with full receipt into stock. The inspection sheet does not state a received-into-stock quantity or a received-into-stock status, so the draft sentence lacks support from the source material.
Configuration and implementation checks
Configuration is built around the checking needs of this page's sample. It lists only settings the official documentation describes, or implementation checks.
| Configuration item | Hypothetical starting value for this trial | Reason and comparison method |
|---|---|---|
| Search mode | Hybrid search | Batch numbers and temperature values need exact hits, and natural-language questions need semantic recall; compare the recalled content for the same batch under the two modes |
| Minimum relevance | Hypothetical starting value 0.3 | This is a 0 to 1 filter threshold that applies only in semantic search or with reranking; adjust after comparing against the hit situation for the batch-number query in the sample |
| Reference limit | Hypothetical starting value 2000 tokens | The official documentation controls the reference limit in tokens; compare whether the inspection sheet and temperature record combined exceed this value |
| Question optimization | Not enabled for now | The sample question contains an explicit batch number; enabling it adds a model call, so compare whether batch-number retrieval results are consistent before and after enabling |
| Rerank | Not enabled for now | A rerank model must be available and enabled per settings; with few candidates in the sample, compare the rerank score ordering against the vector score ordering |
| Chunking method | Chunk by material structure | Temperature record, inspection conclusion, and rejection reason stay in the same data block; compare whether condition and conclusion remain traceable before and after splitting |
Three common errors
- Splitting the temperature record and the inspection conclusion into different data blocks, so a search returns an inspection pass while losing the transport temperature premise. To check, input the chunking result of the sample source material, compare whether the temperature value and the inspection conclusion are in the same data block or have a traceable association, and observe whether searching "batch A-0310 inspection" after splitting returns a fragment missing the temperature premise.
- Writing the draft sentence "has completed inspection and has been fully received into stock" directly into the material as fact. To check, input the draft sentence and the inspection sheet source text, compare whether "sampling passed" and "fully received into stock" are the same fact in the same field, and observe whether the sentence is supported when the source text does not state a received-into-stock quantity or status.
- Treating a retrieval hit as an authorization to execute. To check, input the retrieval request and the handover request, compare the retrieved content against the authorization status field, and observe whether the retrieval result contains states such as approval, release, or settlement that the source text does not state.
How to verify the setup
- Verification one: input the sample source material and the draft sentence "Batch A-0310 has completed inspection and has been fully received into stock." Compare whether the inspection sheet source text states a received-into-stock quantity and status. The source text states only that 6 cartons were sampled and sensory inspection passed; it does not state a received-into-stock quantity or status, so the draft sentence is unsupported and should be recorded as pending verification.
- Verification two: input the batch number "A-0310" and the temperature value "2 to 6 degrees Celsius" as separate searches. Compare whether the data blocks returned by the two search methods contain the full context of the temperature record and the inspection conclusion. If a returned fragment contains only the inspection conclusion and lacks the temperature premise, the chunking method needs adjustment.
- Verification three: input an empty-result scenario, with a paper input of a batch number that does not exist. Compare whether the returned result is empty, and check whether the downstream design handles it as an empty result. An empty result only indicates that this search term did not match; it does not indicate that the material does not exist or was not trained.
Retrieval and relevance do not carry permission or entity isolation functions. The same-vector-model requirement is limited to this knowledge base and does not extend to other systems. After material changes, the actual training and index status must be checked; a successful upload does not mean retrieval has been updated.