What the documents look like
Wholesale stall trading runs on short, scattered paperwork: purchase notes, handwritten sales slips, stall lease contracts and payment receipts. Updates arrive daily or per transaction, and the trade itself leans on on-site bargaining and consignment shipping. A purchase note from one supplier may carry only the item name, piece count, unit price and date. A handwritten sales slip often abbreviates the buyer's name. The vector model and index step has to turn these short texts into retrievable vectors, keep one consistent vector model inside a knowledge base, and keep building indexes as new slips arrive so retrieval can reach same-day or recent documents.
Two features shape this page. First, the association rule: knowledge bases selected together must use the same vector model, so the actual model and its input types need checking before association. Second, one original data block can map to several indexes, and the index content is checked separately from the original text. Index content built from complete, focused information improves precision, while retrieval coverage and index length trade off against each other and need to be weighed against samples. Image retrieval additionally depends on whether descriptions or image vectors were generated at ingestion time, not only on what the query side sends.
Hypothetical paper sample for review. Material text A: "On 12 March 2025, Chen's stall purchased 30 cases of walnut kernels from Xinglong Dried Fruit, at 280 yuan per case, receipt number A-0312; the consignment note is still to be supplied." Material text B: "The stall lease contract shows Chen's stall leased from 1 January 2025 to 31 December 2025, monthly rent 4500 yuan." Draft sentence to be judged: "Chen's stall has settled all payments owed to Xinglong Dried Fruit." The draft concerns settlement status. The source text gives purchase quantity, unit price and receipt number only; it does not state that payment was completed or settled. The draft is unsupported.
Constraints at the vector indexes step
Retrieval recalls candidate content. It does not carry authorization or entity separation. The minimum relevance setting is a filter threshold, and semantic recall together with a relevance score does not exclude unrelated businesses or unauthorized records. A pending request is not approval, and an offered option is not execution.
Knowledge bases associated for retrieval must share one vector model. Check the actual model identifier and its input types first. A text vector model and an image-capable vector model cannot be associated together; identify the mismatch, exclude the non-conforming base and record what was observed, without running a cross-model retrieval to compare missed recalls.
Index content and original content are verified field by field, not merged into one check. Where a model or index changes, compare cited original text against the same question set rather than treating a vector score as a promise of quality. Image retrieval requires a corresponding image input and model capability, and ReRank requires a model enabled under its settings. RRF is a fusion process for mixed results and is not written as a switch. Forms collect values; business verification, approval, database writes and external synchronization each need their own downstream implementation.
Configuration and implementation checks
The table below covers settings and implementation checks tied to this page's documents and task. Values are hypothetical starting points for this paper exercise, not official defaults.
| Configuration item | Hypothetical starting point | Reason and comparison method |
|---|---|---|
| Vector model consistency | Associate one knowledge base and record its vector model identifier | Compare against the model identifier of the slips to be ingested; if they differ, do not associate them together |
| Index content | Build each purchase note's index from item name, quantity, unit price, date and receipt number | Compare field by field against the original text and check whether the subject or validity scope is missing |
| Original data and indexes | One original data block maps to one group of indexes | Verify that index fields and original content are checked separately, not merged |
| Search mode | Semantic retrieval and full-text retrieval used together | Compare full-text hits on number-type queries and semantic hits on natural-language queries |
| Minimum relevance | Hypothetical starting point 0.3 for this exercise | Compare cited original text against the same question set and observe whether the target slip survives filtering |
| Citation limit | Hypothetical starting point 2000 tokens for this exercise | Compare against the actual chunk lengths in that base and observe whether cited content is over-trimmed |
Three common errors
- Treating retrieval output as authorization evidence. Input the draft sentence "Chen's stall has settled all payments owed to Xinglong Dried Fruit" and compare the cited original text returned by retrieval against whether it contains a settlement, payment-completed or receipt-voided field. If the original text holds only purchase quantity, unit price and receipt number, a retrieval hit does not make the settlement fact hold, and the draft stays unsupported. Executable check: record the original sentence returned, compare field by field for settlement, payment-completed or similar status words, and mark the item as pending verification when none appear.
- Associating knowledge bases across vector models. Input two knowledge base identifiers and compare whether their vector model identifiers match. If one is a text vector model and the other an image-capable vector model, they cannot be associated together, and no cross-model retrieval is run to compare missed recalls. Executable check: read the model identifier field of both bases, exclude the non-conforming one when they differ, and record the actual observation without inferring runtime results.
- Replacing field checks with similarity scores. Input a natural-language question and compare the vector score of returned content against the subject, date and amount fields in the original text. A high score does not mean the subject or period matches. Executable check: for each cited item returned, verify subject name, date range and amount fields one by one against the object the question concerns, and mark it as not satisfied when they differ.
How to verify the setup
- Verification one: input material text A and the draft sentence "Chen's stall has settled all payments owed to Xinglong Dried Fruit", then compare whether the cited original text returned by retrieval contains settlement, payment-completed or receipt-voided fields. The expected observation is that the cited original text contains only purchase quantity, unit price and receipt number, with no settlement status word, so the draft is judged unsupported. This verification does not claim to have been run.
- Verification two: input the stall lease contract text and the draft sentence "Chen's stall lease covers the whole of 2025", then compare the contract's start and end dates against the period stated in the draft. The expected observation is that the start and end dates match the draft and the lease coverage holds; if the contract does not state start and end dates, record the item as unsupported. This verification does not claim to have been run.
- Verification three: input an empty-result continuation case, using the explicit paper input [], and check the downstream design against it, comparing whether the downstream chain still runs and outputs a citation array of length zero. The expected observation is that the output chain runs through with an empty citation array. This does not claim that a fictional rule retrieval necessarily returns []. This verification does not claim to have been run.
These checks cover only the input, comparison and observed result of the vector model and index step. Form value collection, business verification, approval, database writes and external synchronization need their own downstream implementation, and neither retrieval nor a form constitutes authorization to execute. Image retrieval requires a corresponding image input and model capability; ReRank requires a model enabled under its settings; RRF is a mixed-result fusion process.