What the documents look like
Pinduoduo low-price campaign operations refresh records daily and per campaign. Four document families matter for this page: campaign registration forms, product cost worksheets, after-sales refund slips, and platform in-site messages. They arrive from the merchant backend, campaign resource-slot data, ERP inventory, and logistics delivery records. Merchants, campaign operators, and supply-chain buyers use them to check registration conditions, cost definitions, and refund status. A retrieval citation has to keep the campaign session, the date, and the version attached to the passage it returns.
The brief marks this category as roughly described rather than confirmed. Treat the update rhythm as daily and per campaign, and treat the document list as the working set for comparison rather than a complete inventory. Nothing here establishes that any other category lacks online certifications or electronic contracts.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Libraries built on different vector models cannot be associated together. Before a run, check the vector model identifier of each library, exclude any that do not match, and record what was observed. Do not run a cross-model search in order to compare missed recalls.
Search mode is chosen per test question. Semantic search, full-text search, and hybrid search are compared against the same question rather than assumed to suit every document type. The citation limit is measured in tokens, so material of different lengths is needed to check whether returned content stays complete. The minimum relevance value takes effect only under semantic search or when result reranking is used; label relevant and irrelevant samples before adjusting the threshold. Question optimization performs coreference resolution and question expansion, so the original question and the optimized question are compared to check whether the subject was altered.
FastGPT uses the PostgresSQL PG Vector plugin as the vector retriever, with an HNSW index. The MongoDB dataset.datas table stores the original vector data together with an indexes field. During retrieval, vectors are recalled first, then the original content is located in MongoDB by vector ID; records that map to the same source data are merged and the highest vector score is kept. RRF is the fusion step applied after hybrid retrieval and is not a separate switch. Reranking requires a configured model and must be enabled in settings. Image retrieval requires an image input and the corresponding model capability. The output reference content is returned as an array and may have a length of zero, so the downstream path still runs when nothing is found.
Retrieval and recall do not perform authorization judgment. Which materials, operations, or requests sit in which authorization state during an operations handover depends on the source text and manual checking. A retrieval result or a form that collects values does not constitute an authorization action.
Hypothetical paper exercise: a Pinduoduo campaign operations team compares contract and transaction materials for one low-price campaign during the period 1 June 2025 to 30 June 2025. The source text supplied for checking reads: "The campaign registration form states that the registration deadline for this campaign is 10 June 2025 at 18:00, the registered product is a summer short-sleeve T-shirt, the campaign price is 19.9 yuan, and the registered stock is 500 units. The product cost worksheet states that the unit cost of this product is 14.5 yuan, the logistics cost is 2.8 yuan, and the gross margin at the campaign price is 2.6 yuan. The after-sales refund slip states that a refund slip dated 15 June has the status pending review, and the refund reason is recorded as incorrect size." The draft sentence to be judged reads: "The summer short-sleeve T-shirt registered for this campaign has passed platform review, the campaign stock of 500 units has been confirmed, and the refund slip has been fully processed." The task is to check whether the three states "has passed platform review," "has been confirmed," and "has been fully processed" agree with the source text.
Configuration and implementation checks
| Configuration item | Starting point for this hypothetical exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Three libraries sharing one vector model identifier: campaign registration forms, cost worksheets, and after-sales refund slips | Check that the three identifiers match before associating them together; exclude any mismatched library one by one and record the observation |
| Search mode | Run semantic search and full-text search once each, then run hybrid search | Compare recalled content across the three modes for the same test question; do not assume one mode suits every document type |
| Citation limit | Starting point for this hypothetical exercise set to 3000 tokens | Test with real material of different lengths to see whether content stays complete, then compare against shorter and longer values |
| Minimum relevance | Starting point for this hypothetical exercise set to 0.5 | Label relevant and irrelevant samples first, then adjust the threshold; the value takes effect only under semantic search or result reranking |
| Question optimization | Enabled; compare the original question with the optimized question | Check whether the subject was altered; coreference resolution and question expansion add one model call |
| Result reranking | Enabled per settings; inspect the rerank score | Reranking produces a 0-1 score that can be used for filtering; confirm the rerank model is configured |
Three common errors
- Treating a retrieval result as the basis for an authorization action. When a refund slip is in pending review status, retrieving that slip does not mean the refund was approved or executed. The input is the source text of the pending-review refund slip, the comparison object is the phrase "has been fully processed" in the draft sentence, and the observable result is that the two states do not agree, so the item is recorded as pending verification.
- Associating libraries that use different vector models. The input is the vector model identifier of each of the three libraries, the comparison object is the association requirement of the node, and the observable result is that mismatched libraries are excluded and recorded. Do not run a cross-model search in order to compare missed recalls.
- Treating RRF fusion as a separate switch that can be turned on. The input is the hybrid retrieval parameters, the comparison object is the description of RRF in the documentation, and the observable result is that RRF is the fusion step applied after hybrid retrieval and is not a standalone configuration item.
How to verify the setup
- Input the source text and the draft sentence from the hypothetical exercise, compare the three states "has passed platform review," "has been confirmed," and "has been fully processed" against the source text, and observe that none of the three is supported by the source text, so each is recorded as pending verification. No test is claimed to have been performed.
- Input the same test question and run semantic search, full-text search, and hybrid search separately. Compare the reference content arrays returned by the three modes and record the actual citations from each mode and whether any difference appears, without assuming a difference must occur. For an empty result, check the downstream design against an explicit paper input of [] and do not claim that a fictional rule search necessarily returns [].
- Input the vector model identifiers of the three libraries and compare them against the association requirement of the node. The observable result is that libraries with matching identifiers can be associated together, while mismatched ones are excluded and the actual observation is recorded. No guarantee is made that the first hit will be correct, that all relevant content will be recalled, or that an unrelated query will necessarily return nothing.