What the documents look like
Overseas third-party warehouse replenishment serves several sales channels at once, and the seller decides how stock is allocated across them. The working documents therefore span four families: replenishment plans, inbound booking requests, inventory age tables, and last-mile delivery orders. These come out of the overseas warehouse WMS, ERP sales forecasts, freight forwarder systems, and carrier statements. Weekly and daily update rhythms coexist, and the same shipment may carry different identifiers, dates, and status wording in different systems. Multilingual retrieval has to handle exactly this cross-system, cross-language comparison: a Chinese question needs to reach English or local-language replenishment clauses, booking windows, and delivery requirements.
FastGPT builds a knowledge base from library, collection, and data. The smallest search unit is the library; collections only classify data and do not affect search results. Vector retrieval uses the PostgresSQL PG Vector plugin with an HNSW index, while MongoDB stores source data. In the MongoDB dataset.datas table, an indexes field records the corresponding vector IDs as an array, so one data item can map to several vectors. A single replenishment clause can therefore carry multiple indexes, and a hit on any one of them recalls that data, with the vector score taking the highest value.
Constraints at the vector indexes step
Knowledge bases linked for retrieval must use the same vector model; the knowledge base search node can only associate libraries that share one vector model. Semantic retrieval carries cross-language understanding, so a Chinese question can match an English knowledge point. That ability comes from the vector model itself and has no connection to permission filtering. The minimum relevance value is a 0-1 filter threshold that applies only under semantic retrieval or when result reranking is used; it filters low-relevance results and does not perform entity isolation or authorization decisions. Hybrid retrieval merges vector and full-text result sets through the RRF formula, and RRF is a fusion process rather than a separate switch. On PG and similar vector stores, full-text retrieval runs through MongoDB $text with jieba tokenization, which suits exact queries for identifiers, model numbers, and proper nouns. The citation limit is controlled in tokens rather than top k.
Hypothetical paper exercise sample
The following material was constructed for this paper exercise and does not correspond to any real customer or real document.
Source text one: "Overseas warehouse A, 2025-03 replenishment plan, clause 7: the inbound booking window for SKU-4471 is 18 March to 22 March; the booking status is awaiting overseas warehouse confirmation; the last-mile delivery order has not yet been generated."
Source text two: "Forwarder B quotation 2025-03-05: the full-container sea freight quotation for SKU-4471 is valid until 31 March 2025; space is subject to actual booking confirmation."
Draft sentence to be judged: "SKU-4471 has completed the inbound booking, the last-mile delivery order has been generated, and the supply period covers the whole of March 2025."
Judging points: source one states the booking status is awaiting overseas warehouse confirmation and the delivery order has not been generated; source two states only a quotation validity date and no supply period. All three statements in the draft sentence lack supporting evidence from the same subject, same period, and same field.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Link only two libraries that use the same vector model | Check the actual model identifier and input type first, exclude non-matching models, and do not run cross-model retrieval comparisons |
| Search mode | Hybrid retrieval | Replenishment plans contain natural-language clauses while bookings and delivery orders contain identifiers; the two recall paths are merged by RRF |
| Minimum relevance | 0.35 | A starting point for this exercise; compare the same set of replenishment questions at 0.35 and 0.5 to see whether cited text still contains unrelated fragments |
| Citation limit | 3000 tokens | A starting point for this exercise; observe whether the same question at 3000 and 5000 tokens covers the clauses required |
| Result reranking | Enabled | Used when candidate results are numerous and the question has complete subject and predicate; the 0-1 rerank score can be used for filtering |
| Index content | Build a separate index for each replenishment clause | Complete and focused indexes raise precision; index length and recall range need to be weighed against the sample |
Three common errors
- Treating semantic recall as an authorization decision. A retrieved replenishment plan only means the content is semantically close to the question; it does not mean the current operator is entitled to execute an inbound booking or change stock allocation. Authorization status must be confirmed separately in the business system.
- Reading "awaiting overseas warehouse confirmation" as a completed booking. Source one gives the status field as awaiting overseas warehouse confirmation and states the delivery order has not been generated, while the draft sentence writes completed booking and generated delivery order; these are mutually exclusive values in the same field.
- Treating quotation validity as a supply period. Source two states only that the quotation is valid until 31 March 2025 and gives no supply period, so the draft sentence's claim that the supply period covers the whole month lacks supporting evidence and should be recorded as pending verification.
How to verify the setup
- Input source text one together with the draft sentence, and compare the three statements in the draft against the status field and the delivery-order generation state in the source. The source states awaiting confirmation and no delivery order; the draft states completed and generated. A mutually exclusive value in the same field means the draft sentence does not hold.
- Input source text two together with the supply-period statement in the draft, and compare the quotation validity field with the supply period field. The source carries only a quotation validity date and no supply period, so the statement is judged unsupported; keep the evidence gap open and do not supply authorization or effective clauses.
- Input a query that is clearly not contained in the linked knowledge bases and observe the returned citation content. Citations are output as an array whose length may be 0. If an empty array is returned, use the explicit paper input
[]to check whether the downstream chain handles it normally, without claiming that this query necessarily returns an empty result.
This page addresses only overseas third-party warehouse replenishment, cross-border material and multilingual retrieval, and vector models and indexes. Which materials, operations, and requests sit at which authorization state during an operations handover cannot be settled by retrieval or by a form.