What the documents look like
Selection and sample inspection happen before a livestream opens. The work decides whether a product can be sold and whether it suits on-camera display. Source material arrives as supplier files, testing-agency reports, platform negative lists, and historical return rates. Day-to-day documents include sample evaluation forms, quality inspection reports, supplier qualification files, and prohibited-sale lists. These materials update weekly or per item; the update rhythm itself does not change the retrieval configuration on this page.
For vector models and indexes, document shape drives three decisions. Sample evaluation forms and inspection reports often carry precise fields such as item number, model, and tested parameter, so full-text retrieval and semantic retrieval serve different purposes. Supplier qualification files and prohibited-sale lists usually appear as entries, and one original data block can carry several indexes. Whether image material produced an image description or an image vector at ingestion decides whether a later image query can match; uploading an image at query time is not sufficient on its own.
Constraints at the vector indexes step
Knowledge bases linked for retrieval must share the same vector model. Check the actual model identifier and input type first, exclude any library that does not match, and do not run cross-model retrieval to compare recall differences. One original data block can be associated with multiple indexes; verify index content and original content separately. Indexes should use complete and focused information, and retrieval coverage against index length needs to be weighed against the sample. Image retrieval also depends on whether an image description or image vector was generated at ingestion.
FastGPT uses the PostgresSQL PG Vector plugin as the vector retriever, with HNSW as the index, while MongoDB handles other data storage. The dataset.datas table in MongoDB stores original vector data information, and an indexes field records the corresponding vector IDs as an array. During retrieval, vectors are recalled first, then the vector ID is used to find original content in MongoDB; entries mapping to the same original data are merged, and the highest vector score is kept. Hybrid retrieval merges multiple result paths through RRF, which is a fusion process rather than a switch that can be turned on. The reference limit is controlled by tokens. Minimum relevance takes effect only under semantic retrieval or when result reranking is used; it is a filter threshold and does not perform permission or entity isolation. A retrieval action does not itself constitute authorization.
Configuration and implementation checks
The table below lists officially documented settings or implementation checks tied to this page's primary task. The values are starting points for this hypothetical exercise, not official defaults.
| Configuration or check | Hypothetical starting point and reason | Comparison method and observation target |
|---|---|---|
| Vector model of linked knowledge bases | Register the same model identifier across all candidate libraries; verify before linking | Compare model identifier and input type library by library; exclude mismatches |
| Index content | Build indexes for the sample evaluation form segmented by tested parameter | Use the same question set to compare whether cited original text covers that parameter |
| Original data and index mapping | Attach multiple indexes to one original data block; list index and original text separately | Check whether the indexes array matches the original content |
| Image index | Whether ingestion generated a description or image vector for the image | Inspect the ingestion record, checked separately from the query-time upload action |
| Reference limit | Set to 2000 tokens for this exercise, because chunk lengths differ widely in mixed libraries | Compare against the total cited content returned for the same question set |
| Minimum relevance | Set to 0.3 for this exercise, to observe remaining entries after filtering | Compare entries before and after filtering, only under semantic retrieval or reranking |
Three common errors
- Treating retrieval or a form as authorized execution. A sample evaluation form collects evaluation values. Whether a supplier qualification is valid, whether a sample is approved, and whether a prohibited-sale list is hit all require downstream business verification or human confirmation. A retrieved qualification record does not mean the supplier passed review.
- Describing RRF as a switch, or treating minimum relevance as a permission filter. RRF is the fusion process for hybrid retrieval results. Minimum relevance is a 0–1 filter threshold and does not exclude unrelated businesses or unauthorized records.
- Inferring that an image can be retrieved from a query-time upload. Whether an image can be retrieved also depends on whether a corresponding index was created at ingestion, and whether a vision model or an image-capable vector model is configured.
How to verify the setup
The following uses one hypothetical exercise sample. Sample material text: a sample evaluation form submitted by a supplier states sample number YP-2024-031, tested parameter fabric composition, and evaluation status pending review; an attached inspection report states the same number and a conclusion of compliant with the commissioned item; a supplier qualification file states the entity as a certain trading firm, valid until 31 December 2024. Proposed handover request: add this sample to this week's livestream selection candidates and mark the qualification as passed.
- Input the proposed handover request and the three material texts; compare "qualification passed" in the request against the validity period stated in the qualification file and the evaluation status "pending review." If the materials contain no record of a passed review, the statement is unsupported, recorded as pending verification, and not judged as a denial.
- Input the same question set to retrieve this sample number and tested parameter; compare whether the returned cited original text contains YP-2024-031 and fabric composition. If the cited original text covers that number and parameter, the index content corresponds to the original text; if it does not cover them, record the gap and do not infer a missed recall.
- Input a query containing only an image; compare whether the ingestion record shows a description or image vector generated for that image. If the ingestion record states no corresponding index, the image retrieval condition is not met, the result is unknown, and no claim is made that an empty result must be returned. If downstream accepts an empty array, a clearly labelled paper input of [] can be used to check the downstream design, without claiming that a fictional rule retrieval necessarily returns [].
This page addresses only livestream selection and sample inspection / product material and selection questions / vector models and indexes. When operations hand over materials, actions, or requests, their authorization state must be checked; retrieval or a form must not be treated as authorized execution.