What the documents look like
Marketplace promotion registration and price management revolves around price, inventory and campaign cadence. The data sources named in the brief are the Tmall marketing backend, the JD campaign backend, ERP inventory and historical transaction prices. For the inventory procurement and supplier materials task, the everyday documents are campaign registration forms, promotional price approval sheets, inventory lock sheets and campaign review sheets. The brief gives the update frequency as weekly and per campaign.
These materials carry two kinds of content at the same time. They contain identifiers that are matched exactly, such as supplier batch numbers, SKUs and approval sheet numbers. They also contain condition descriptions that need semantic understanding, such as the effective scope of an inventory lock and the period in which a price constraint applies. Document parsing and chunking has to serve both retrieval paths. Identifiers are found through full-text retrieval, and condition descriptions are found through semantic retrieval. When a document is split apart, a condition separates from the object it constrains, and a retrieved citation can arrive without its subject or its validity period.
Constraints at the document chunking step
Chunking follows the structure of the material itself, so that a condition, its explanation and the fields it constrains stay inside the same data group. In FastGPT the knowledge base consists of libraries, collections and data. The smallest search unit is the library; collections only classify data and do not affect search results. One data group can map to multiple vector indexes, and the indexes field records the corresponding vector IDs. During retrieval, vectors are recalled first, then the vector ID is used to fetch the original data from dataset.datas in MongoDB. Records that map to the same original data group are merged, and the vector score takes the highest value. A longer text can be given multiple indexes to raise precision, at the cost of a trade-off between content length and semantic richness.
Search modes are semantic retrieval, full-text retrieval and hybrid retrieval. Hybrid retrieval combines multiple result paths through RRF. A ReRank model produces a 0-1 score that can be used for filtering. The minimum relevance value takes effect only under semantic retrieval or when result reranking is enabled, and it is a filter threshold rather than a permission or entity isolation mechanism.
Hypothetical paper exercise sample: a brand sales planning role hands over a paper promotional price approval sheet. The three source sentences are: "Approval sheet number PA-2026-0417. Applies to SKU-8801 and SKU-8802 supplied by East China warehouse Class A suppliers. Locked price validity 1 May 2026 to 7 May 2026; during this period the locked price must not exceed 199." The same handover contains a request written as: "Please lock inventory for the whole of May according to this sheet and synchronize it to the JD campaign backend." That request extends the locked price scope from 1 May to 7 May to the whole month, and it includes an action to synchronize to an external system. Both elements exceed the authorization stated in the source text.
Configuration and implementation checks
| Configuration or check item | Setting for this page | How to check and what it means |
|---|---|---|
| Search mode | Semantic retrieval, full-text retrieval, hybrid retrieval | Use full-text or hybrid retrieval for identifiers and approval sheet numbers; use semantic or hybrid retrieval for locked price condition descriptions |
| Result reranking | ReRank second-pass ordering | Enable when candidates are numerous and the question is complete; filter by the rerank score |
| Citation limit | A limit measured in tokens | In mixed knowledge bases chunk lengths vary widely, so citation volume is controlled in tokens |
| Minimum relevance | A 0-1 threshold | Takes effect only under semantic retrieval or when reranking is enabled; adjust against observed results |
| Index and segmentation | One data group can map to multiple vector indexes | Split according to material structure and keep a condition and the fields it constrains in the same data group |
| Training and update check | Check upload and actual training and index state separately | A completed upload does not mean retrieval has updated; re-check the actual state after a change |
Hypothetical starting value for this exercise: the citation limit is first set to a small value that can hold the approval sheet number, the applicable object and the locked price period across three sentences. The reason for choosing it is that the sample source text has only three sentences, which makes it easy to observe whether the condition is recalled in the same group as its object. The comparison target is whether the citation content retrieved before and after splitting the same material remains complete. This starting value is hypothetical and is not an official default.
Three common errors
- Splitting a condition and its object into different chunks. If "Locked price validity 1 May 2026 to 7 May 2026" is separated from "SKU-8801 and SKU-8802" in the sample, a retrieval hit on the validity period returns a citation that lacks the applicable object.
- Treating retrieval or a form as authorized execution. The handover request asks to "lock inventory for the whole of May according to this sheet and synchronize it to the JD campaign backend." The source text covers only 1 May to 7 May and states no authorization for synchronization. That request remains unsupported.
- Treating a successful upload as an updated retrieval state. After a material change, the actual training and index state has to be checked. A completed upload alone does not confirm that the new segmentation can be recalled.
How to verify the setup
- Input the three source sentences and the search term "PA-2026-0417", then compare whether the recalled citation contains the approval sheet number, the applicable object and the locked price period together. If the citation contains only the number and lacks the locked price period, segmentation or indexing has not preserved the condition and object in the same data group.
- Input the proposed handover request "Please lock inventory for the whole of May according to this sheet and synchronize it to the JD campaign backend", then compare the request scope against the source locked price scope. The source text covers only 1 May to 7 May and states no external synchronization authorization. The request exceeds authorization and should be recorded as pending verification with the evidence gap preserved.
- Input the paper empty-result sample
[]and check how the downstream design handles no recall, confirming that an empty result is not treated as authorized or executed. This check does not claim that a fictional rule search necessarily returns an empty array. - For a library that spans different vector models, check the model identifier first and exclude those that do not match, recording the actual observation. Do not run a cross-model retrieval to compare missed recalls. Image retrieval and reranking are optional paths that require the corresponding image input and model capability and must be enabled according to the settings; they do not mean every request executes all paths.