What the documents look like
Receiving and put-away covers the first physical-and-record confirmation that goods have entered the warehouse. Data comes from the WMS, ERP purchasing, supplier delivery notes and barcode scanning. The documents handled each batch are purchase arrival notes, receiving inspection sheets, put-away tasks and discrepancy reports.
In a cross-border setting, one arrival may carry a Chinese purchase order, an English packing list and destination-language labels at the same time. Document parsing and chunking must preserve a document identity and version that can point back to the original text, so multilingual retrieval can check against it. Verifying arrived quantities, printing barcodes and assigning storage locations are the time-consuming steps. If a chunk separates a condition from the content it constrains, the retrieved fragment loses its subject or its effective scope.
Constraints at the document chunking step
Chunk completeness takes priority over fragment length. A longer passage can be mapped to several vectors, and a hit on any one of them recalls the underlying data group; annotating data blocks can raise precision. Original data is stored in the MongoDB dataset.datas table, whose indexes field records the corresponding vector IDs, so one data group can correspond to multiple vectors. PostgresSQL stores vectors in a vector field with an HNSW index. Retrieval recalls vectors first, then uses the vector ID to fetch the original data from MongoDB; records that map to the same data group are merged and the highest score is kept.
Semantic search supports cross-language understanding, while full-text search suits numbers, models and proper nouns. Hybrid search merges results through RRF, and a rerank score can be used for filtering. The minimum relevance value takes effect only under semantic search or result reranking, and it is a filtering threshold; it does not perform permission or entity isolation. A similarity score is not a category, entity or permission filter, and retrieval is not authorization.
Configuration and implementation checks
The following paper exercise is used only for desk checking. A hypothetical warehouse receives a supplier delivery note whose original text reads: "Delivery note SD-2024-0871, purchase order PO-2024-3312, arrival date 3 June 2024, material M8 bolt, quantity 1200 pieces, status is pending inspection." The sentence under judgment is: "This batch of M8 bolts has completed put-away." The handover request to be assessed is: "Please arrange put-away per this delivery note and write back to inventory." This sample is fictional and no business action has occurred.
| Check item | Hypothetical starting point for this exercise | Reason and comparison object |
|---|---|---|
| Chunk boundary | Split at document number and material line | Compared with leaving the passage undivided, to check whether the subject survives in the fragment |
| Index count | One vector per chunk | Compared with multi-vector annotation, to observe recall coverage |
| Search mode | Semantic search | Compared with full-text search on number-type queries |
| Minimum relevance | No uniform value set | Compared with filtering when semantic search or reranking is enabled |
| Reference limit | Measured in tokens | Compared with a top-k measure, to observe chunk length differences |
| Update recheck | Verify training and index state after upload | Compared with checking upload success alone |
The reference limit is expressed in the documented unit of tokens. Same-model retrieval association, empty-reference output and context limits keep their documented conditions. Forms collect values; business verification, approval, database writes and external synchronization require their stated downstream implementation.
Three common errors
- Treating a successful upload as an updated index. Input: the delivery note upload record. Comparison: the actual training and index state. Observable result: retrieval can return that original text; upload success alone does not satisfy this.
- Treating a retrieval hit as authorization. Input: the handover request under assessment. Comparison: whether put-away authorization and write-back permission exist. Observable result: the authorization state is explicit; a similarity score does not constitute authorization.
- Treating pending inspection as completed put-away. Input: the status field in the sample original text. Comparison: the sentence under judgment. Observable result: a contradiction requires mutually exclusive values in the same field for the same subject and the same period; here the status is pending inspection, and the sentence under judgment is unsupported.
How to verify the setup
- Verification one: input the sample original text and the sentence under judgment, compare the values of the status field, and observe whether mutually exclusive values appear in the same field. If they do not appear, the sentence under judgment remains unverified.
- Verification two: input delivery note number SD-2024-0871 as a full-text search, compare whether the returned fragment contains that number and the material line, and observe whether the subject and effective scope are preserved.
- Verification three: input an empty result
[], compare whether the downstream design distinguishes an empty set from a search that was never run, and observe whether the gap is recorded.
Libraries built on different vector models cannot be associated together; check identifiers first, exclude those that do not match, and record the actual observation. Image retrieval requires a corresponding image input and model capability, and reranking requires a model enabled by its settings. Where no documented error code or actual observation exists, describe what to compare and leave the rest unknown. A pending request is not approval, an offered option is not execution, and an unsigned termination is not a completed termination.