What the documents look like
Multi-warehouse allocation happens at the order routing stage and determines which warehouse ships an order, which distinguishes it from parcel handover. Four document families form the main retrieval surface for operating reviews and exception post-mortems: allocation rules, inventory synchronization tables, transfer orders, and stockout alerts. Source systems include WMS, OMS, ERP inventory, and platform inventory interfaces, with update frequencies of daily or hourly. A review needs to locate, for a given oversell, cross-region freight increase, or shipping delay, the rule basis and the synchronization record behind it. The retrieval result determines which rule or which synchronization table the review conclusion cites.
Allocation rules typically carry warehouse priority and regional constraints. Inventory synchronization tables carry timestamps and sellable quantity fields. Transfer orders carry source warehouse, destination warehouse, and status. Stockout alerts carry trigger conditions. 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. One data item can map to multiple vectors. Retrieval recalls vectors first, then looks up the original data in MongoDB dataset.datas by vector ID; items mapping to the same original data are merged and the highest vector score is kept. Reference content is output as an array and its length can be 0, meaning the output path still runs even when nothing is found. After hybrid retrieval is merged through RRF, similarity filtering cannot be applied directly; reranking is usually needed and filtering uses the rerank score. Minimum relevance takes effect only under semantic retrieval or when result reranking is used, and it is a filter threshold rather than a sorting rule.
A hypothetical paper exercise sample: source text one reads "East China warehouse allocation rule V3, effective 1 March 2025, provides that Shanghai region orders route preferentially to the East China warehouse, and when East China sellable inventory falls below safety stock, shipping transfers to the North China warehouse." Source text two reads "12 March 2025 inventory synchronization table record: East China sellable inventory 1,200 units, North China sellable inventory 800 units, synchronization time 08:00." The draft sentence to be assessed is "On 12 March 2025 all Shanghai orders shipped from the East China warehouse." Whether this draft sentence shares the same subject, the same period, and the same field as source texts one and two must be checked item by item.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Libraries with different vector models cannot be associated together; the identifier is checked first and non-matching libraries are excluded, without running cross-model retrieval to compare missed recalls. A retrieval hit is not an authorization. Recalling an allocation rule or a transfer order only shows the material was recalled; it does not show that the transfer was approved or that the inventory batch was synchronized.
Minimum relevance is a 0-1 filter threshold. It does not exclude unrelated businesses or unauthorized records, and it does not act as a permission or entity isolation mechanism. Similarity settings do not exclude unrelated businesses or unauthorized records. Whether reference content falls within the current user's accessible material scope is a separate access-control question handled by downstream permission implementation.
Question optimization performs coreference resolution and question expansion before retrieval. It adds one model call before formal retrieval, which usually improves multi-turn retrieval but also adds latency. Comparing the original question with the optimized question lets the reviewer check whether the subject was altered incorrectly. RRF is a fusion process for hybrid results, not an enablement switch. Image retrieval and reranking run only when the corresponding input and model capability are present, per the documented conditions.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Allocation rule library, inventory synchronization table library | The two libraries can be associated only if they use the same vector model; check the vector model identifier first and exclude any that does not match |
| Search mode | Hybrid retrieval | Allocation rules contain identifiers and warehouse names, so semantic and full-text retrieval are used together; compare recall differences against semantic retrieval alone |
| Reference limit | Hypothetical starting point 3,000 tokens | Inventory synchronization records are short while allocation rule passages are long; compare content completeness against 1,500 and 6,000 |
| Minimum relevance | Hypothetical starting point 0.35 | Takes effect only under semantic retrieval or result reranking; label relevant and irrelevant samples first, then compare filtering against 0.2 and 0.5 |
| Question optimization | Enabled | Review follow-up questions often contain references; compare the original question with the optimized question and check whether the subject was altered incorrectly |
| Result reranking | Enabled | Hybrid retrieval produces many candidates; compare ordering before and after reranking and confirm the rerank score is used for filtering |
Three common errors
- Treating a retrieval hit as authorized execution. Recalling an allocation rule or a transfer order only shows the material was recalled; it does not show that the transfer was approved or that the inventory batch was synchronized. The input to check is the reference content returned by retrieval, the comparison object is the current status field of that transfer order, and the observable result is whether the status field states approval; when it does not, record the item as pending verification.
- Forcing association of libraries with different vector models. The node requires associated knowledge bases to use the same vector model. The input to check is each library's vector model identifier, the comparison object is the model condition the node allows for association, and the observable result is whether the identifiers match; when they do not, exclude that library and record it, without running cross-model retrieval to compare missed recalls.
- Treating minimum relevance as permission or entity isolation. Minimum relevance is a 0-1 filter threshold and does not exclude unrelated businesses or unauthorized records. The input to check is the reference content returned by retrieval, the comparison object is the material scope the current user can access, and the observable result is whether the reference content falls within the authorized scope; when it exceeds that scope, record it as an access-control gap and hand it to downstream permission implementation.
How to verify the setup
- For the hypothetical sample draft sentence "On 12 March 2025 all Shanghai orders shipped from the East China warehouse," the input is source text one and source text two, the comparison object is the subject, period, and field in the draft sentence, and the observable result is that source text one only specifies preferential routing and a low-inventory transfer condition while source text two only records sellable quantity at 08:00. The word "all" in the draft sentence is not supported by either source text, so the item is assessed as pending verification.
- For the same hypothetical sample, the input is source text one, "when East China sellable inventory falls below safety stock, shipping transfers to the North China warehouse," the comparison object is the East China sellable inventory of 1,200 units in source text two and the safety stock field, and the observable result is that source text two does not state a safety stock value, so whether the transfer condition was triggered cannot be determined and the item is recorded as an evidence gap.
- For empty-result handling, the input is a clearly stated paper reference content of [], the comparison object is the downstream node's design for handling a reference array of length 0, and the observable result is whether the downstream node runs the output path as designed. No claim is made that a fictional rule retrieval necessarily returns [].
Throughout this page, retrieval, forms, and draft sentences are not treated as authorization, approval, or execution results. Fields that are not stated remain unknown. RRF is described as a hybrid result fusion process and is not presented as an enablement switch. Image retrieval and reranking run only when the corresponding input and model capability are present, per the documented conditions.