What the documents look like
A Shopify standalone store lets the merchant configure payment, shipping and fulfillment services directly, and returns are not handled by a single platform-wide automatic process. For inventory procurement and supplier records, the materials that enter the knowledge base are purchase orders, supplier quotations, arrival records and refund evidence. These sit alongside order details, shipping templates, shipping labels and refund records that originate from Shopify Admin, payment gateways, carrier systems and inventory location data. The operational consequence is that procurement evidence and fulfillment evidence are stored in different systems and must be retrieved together when staff answer a question about a purchase.
These materials update per order and daily. Chunking must preserve the identity and version of each document so that a retrieval result can point back to the source text. A supplier quotation carries a quotation number and a quotation validity date; an arrival record carries a received quantity and an inspection status. If a chunk separates a condition from the content it constrains, retrieval can return a fragment that has lost its scope.
Constraints at the document chunking step
In FastGPT a single data item can map to multiple vectors, and the indexes field records the corresponding vector IDs. When several indexes belong to the same original data, the highest score is used after merging. Full-text retrieval depends on keyword hits and suits exact queries such as a purchase order number or a model number. Semantic retrieval depends on vector distance and suits natural-language descriptions. A minimum relevance value is a 0–1 threshold that applies only to semantic retrieval or when result reranking is used; it is a filter threshold and does not perform permission or supplier isolation. Whether an image can be retrieved depends on the model capability and on whether an index was built when the image was stored.
Hypothetical paper exercise sample: the source text reads “Supplier A, quotation number Q-2024-018, quotation valid until 30 June 2024, received quantity 300 units, inspection status pending inspection.” The proposed sentence under review reads “Supplier A has completed inspection of the 300 received units.” The source states only that inspection status is pending inspection and does not state that inspection is complete, so the proposed sentence is unsupported.
Configuration and implementation checks
| Configuration item | Value or check for this page | Explanation and comparison method |
|---|---|---|
| Search mode | Hybrid retrieval | Purchase order numbers rely on full-text hits; descriptive questions rely on semantic recall; results merge through RRF |
| Minimum relevance | Hypothetical trial starting point 0.3 | A trial starting point only; compare against the observed score distribution of returned results before adjusting |
| Reference limit | Set in tokens; hypothetical trial starting point 800 | The unit is tokens; compare against the total length of returned fragments |
| Question optimization | Decide per multi-turn follow-up scenario | Enabling it adds one model call and latency; compare against the response requirement |
| Chunk review | Manually check whether conditions and the content they constrain stay in one chunk | Compare the source structure against the parsed result and keep version traceability |
| Post-update review | Check the actual training and index state | A successful upload is a different fact from retrieval being updated |
Three common errors
- Treating minimum relevance as a supplier or permission filter. The value is a 0–1 similarity threshold that applies only to semantic retrieval or result reranking, and it does not exclude unauthorized records.
- Treating a successful upload as an updated index. After material changes, the actual training and index state must be checked; these are separate facts.
- Writing pending inspection as completed inspection. In the sample the source inspection status is pending inspection, while the proposed sentence states completed inspection, which is unsupported.
How to verify the setup
- Input the sample source text and the proposed sentence, then compare the inspection status field: the source says pending inspection, the proposed sentence says completed, and the observed result is that the sentence is unsupported and the evidence gap must be recorded.
- Input purchase order number Q-2024-018 for full-text retrieval, compare whether the returned fragment points back to that quotation source, and observe whether the number and validity date are preserved; if there is no hit, record the observation without guaranteeing a first-position hit.
- Input a natural-language query for Supplier A quotations, compare the semantic relevance score of returned fragments against the configured threshold, and observe whether the score reaches the threshold. If the result is empty, use the explicit paper input [] to check the downstream design, and do not claim that a rule-based retrieval necessarily returns [].
This page addresses only Shopify standalone order fulfillment / inventory procurement and supplier records / document parsing and chunking. Handover materials, operations and requests carry an authorization state that retrieval and forms do not execute.