Omnichannel stock transfers: document chunking for procurement

What the documents look like

Brand omnichannel stock transfer covers the movement of goods between stores and warehouses. It does not cover marketing outreach. The main data sources are ERP inventory, store POS, e-commerce orders and WMS. Everyday materials include transfer requests, inventory snapshots, store shortage lists and transfer logistics notes. Within the procurement and supplier records task, these materials commonly carry supplier name, purchase batch, arrival quantity, quality inspection conclusion and settlement terms. Those fields cross-reference the store, SKU and quantity held in the transfer request. Errors produce overselling, store stockouts or a mismatch between book and physical stock, so parsing and chunking must keep the constraints in one record together with the content those constraints govern.

Constraints at the document chunking step

A FastGPT knowledge base consists of library, collection and data. The smallest search unit is the library; a collection only classifies data and does not affect search results. One data item can map to multiple vectors, and the indexes field records the array of vector IDs. Retrieval recalls vectors first, then uses the vector IDs to find the source data in the MongoDB dataset.datas table. When several vectors map to the same source data, the results are merged and the highest score is taken. If arrival quantity, quality inspection conclusion and settlement terms in a supplier record are split into separate chunks, a recall can surface only one fragment, and the condition is separated from its conclusion.

Full-text search suits exact queries such as numbers, models, proper nouns and error codes. Semantic search suits natural-language phrasing. Hybrid search uses both recall paths and merges them through RRF. Rerank performs a second ordering of candidate text and returns a score between 0 and 1. The minimum relevance value takes effect only under semantic search or when result reranking is used. It is a filter threshold, not an ordering rule, and it does not perform permission or entity isolation. A retrieval result does not constitute authorization; transfer approval and arrival putaway still require their own downstream implementation.

Hypothetical paper exercise sample

The following is hypothetical paper material. It does not involve a real customer or a real identifier.

Source text (hypothetical): Transfer request T-2024-0312 states that the requesting party is a store in the East China region, the inbound SKU is A-118, the requested quantity is 60 units, and the expected arrival date is 20 March 2024. Supplier record S-207 states that the purchase quantity for this batch is 200 units, the quality inspection conclusion is pass, and the quotation validity period runs to 31 March 2024. Transfer logistics note L-0312 states that 60 units were actually dispatched and the receipt status is unsigned.

Draft sentence to be assessed (hypothetical): Because the quotation validity period in supplier record S-207 runs to 31 March 2024, this supplier can supply 60 units on 20 March 2024.

The draft sentence treats quotation validity as a supply period. These are two different fields, and the material does not state a supply period. The sentence is judged unsupported.

Configuration and implementation checks

The values in the table are hypothetical trial starting points for this page. They are not official defaults.

Three common errors

1. Treating quotation validity as a supply period. Record S-207 states only that the quotation validity period runs to 31 March 2024 and does not state a supply period. The draft sentence infers from this that 60 units can be supplied on 20 March 2024, which substitutes a different field for the basis of the judgment.
2. Treating a successful upload as an updated index. After a material change, if only the upload action is confirmed and the actual training and index state is not checked, retrieval can still return the old chunk.
3. Treating a retrieval hit as transfer authorization. Recalling transfer request T-2024-0312 only shows that the text can be recalled. Approval status and arrival putaway require a separate implementation, and retrieval does not produce authorization.

How to verify the setup

1. Enter request number T-2024-0312 and SKU A-118 from the source text into full-text or hybrid search, and compare whether the returned chunk contains both the requested quantity of 60 units and the expected arrival date of 20 March 2024. The condition is met only when both appear in the same source data item or can be traced back to the same source text.
2. Enter the phrase "quotation validity period runs to 31 March 2024" from the draft sentence into semantic search, and compare whether the returned chunk contains a supply period field. If the returned content shows only the quotation validity period and no supply period, the draft sentence is judged unsupported and the evidence gap is recorded.
3. For the same source data item, check whether indexes contains multiple vector IDs, then search once with a natural-language question and once with a number-based question, and compare whether both recall the same source data item. If the number-based question hits and the natural-language question does not, record this as a chunking or index item to be checked, and do not judge it as missing data.

The above are paper check steps. They were not actually executed and do not constitute a test, deployment or approval conclusion.

Related reading

• guide/dataset/dataset_engine
• FastGPT document chunking