What the documents look like
Amazon FBM fulfillment is operated by the seller, who arranges delivery and sets the return address, without relying on platform warehouse inspection. Inventory procurement and supplier records therefore include purchase orders, supplier confirmations, picking lists, shipment confirmation sheets, return address settings and logistics tracking tables. Sources span Amazon Seller Central, carrier systems, ERP orders and customer service email. These records update per order or per day, and a single mismatch can contribute to order cancellation, A-to-z claims or a drop in store performance.
The physical form of each record shapes parsing. A picking list centers on order number, SKU and quantity. A shipment confirmation sheet carries the carrier and tracking number. A return address setting is configuration-style text. A logistics tracking table is a row-per-order status list. During chunking, the constraint relationships among order number, supplier, quantity, date and status must stay together, and conditional statements cannot be separated from the content they govern.
Constraints at the document chunking step
In FastGPT, a knowledge base consists of library, collection and data. The smallest search unit is the library; collections only classify data and do not affect search results. The MongoDB dataset.datas table stores source data, and its indexes field records vector IDs, so one data group can map to multiple vectors. The PostgresSQL vector field stores vectors. Retrieval recalls vectors first, then looks up source content in MongoDB by vector ID; content matching the same data group is merged, and the highest vector score is taken. Multi-vector mapping balances content length against semantic richness, letting one chunk be hit through several indexes.
Semantic search relies on vector distance, full-text search relies on keyword hits, hybrid search merges both through RRF, and ReRank reorders candidates and returns a 0-1 score. The citation limit is measured in tokens. Minimum relevance takes effect only under semantic search or result reordering. Image retrieval requires image input and matching model capability; reranking requires a model enabled per settings. Optional paths shown in the overall flow diagram are not all executed on every request. Retrieval and similarity do not perform permission or entity isolation, and libraries using different vector models cannot be associated together.
Configuration and implementation checks
The following items concern a hypothetical drill sample and list only documented settings or implementation checks tied to this feature and the inventory procurement and supplier records task. Values are starting points for this hypothetical trial, not official defaults.
| Item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| Search mode | Hybrid search | Purchase order numbers and supplier names need exact hits, and natural-language questions also occur; compare against pure semantic search for whether order numbers are covered |
| Minimum relevance | 0.5 | This value takes effect only under semantic search or result reordering; compare with 0.3 and 0.7 for whether filtered fragments contain conditional statements |
| Citation limit | Measured in tokens, hypothetical starting point 2000 | Chunk lengths vary widely in mixed knowledge bases and top k is unstable; compare with 1500 for whether returned content truncates quantity constraints |
| Question optimization | Enabled for multi-turn follow-up scenarios | Coreference resolution and question expansion add one model call and latency; compare with disabled for short-question recall |
| Result reordering | Enabled per settings | The 0-1 rerank score can be used for filtering; compare with no reranking for candidate ordering changes |
| Image index | Confirm at ingestion whether generated | Whether an image can be retrieved depends on the ingestion index and model capability; check actual index status |
Three common errors
- Treating a successful upload as retrieval already updated. After a record changes, check actual training and index status; the upload action and index taking effect are two separate facts.
- Treating a similarity score as a permission or entity filter. Minimum relevance is a filtering threshold and does not exclude unrelated businesses or unauthorized records; retrieval does not constitute authorization.
- Treating a pending review request as approved and executed. Pending review does not demonstrate lack of operating permission; signing, reaching cooperation and contract effectiveness are three distinct facts, and a quote validity period does not indicate a supply period.
How to verify the setup
Hypothetical drill sample: the source text reads "Purchase order PO-2201, supplier is a packaging materials factory in East China, purchasing 2000 corrugated cartons, agreed delivery date March 10, 2025. The supplier confirmed acceptance of the order and has not yet signed the annual framework contract." The draft sentence to be judged is "The supplier has signed the annual framework contract and payment can be arranged."
- Input the source text above and the draft sentence, then compare whether "confirmed acceptance, not yet signed" and "has signed" refer to the same subject, same period and same field. The source supports only "confirmed acceptance"; the draft sentence's "has signed" is unsupported and remains to be verified.
- Input purchase order number "PO-2201" and the supplier name, then use full-text search and semantic search separately to observe whether returned fragments contain the quantity and delivery date. If conditional statements and quantity are split apart, adjust the chunking and recheck.
- Input the paper empty result [] to check downstream design: when retrieval returns nothing, check whether the flow keeps a gap record, and do not interpret the empty result as the supplier lacking qualification or the order not existing.