What the documents look like
Marketplace advertising and on-site traffic operations draw on sources such as Juliang Qianchuan, JD Kuaiche, Alimama and store transaction data. The working document set includes campaign plans, keyword reports, audience package descriptions and conversion attribution tables, with updates occurring daily or hourly. In this category, contract and transaction material comparison means checking consistency of subject, period and field across advertising service contracts, agency operation agreements, settlement statements and attribution tables.
Document parsing and chunking determine whether these materials can be retrieved at the correct passage. If the authorization scope in an advertising contract, the billing period in a settlement statement and the conversion definition in an attribution table are split across different chunks, a comparison cannot be completed within a single citation. This category adjusts prices in real time according to traffic and conversion, and material versions change with the campaign cycle, so chunking must preserve document identity and version that can be traced back to the source text.
Constraints at the document chunking step
Chunking must follow material structure. Conditions, explanations and the content they govern cannot be separated. The authorization boundary in an advertising contract, the quantity and unit price in a settlement statement, and the statistical period in an attribution table belong to the same judgment unit; once split, any single chunk cited alone is insufficient to support a comparison conclusion.
FastGPT uses multi-vector mapping, where one set of source data can correspond to multiple indexes. Retrieval recalls that data when any one index is matched, and the vector score takes the highest score. This requires that multiple indexes of the same chunk remain semantically consistent, and that indexes of different fields are not mixed into the same set of source data. The citation limit is measured in tokens. In a mixed knowledge base, chunk lengths vary widely, so chunking granularity directly affects the range of material that can be cited in a single retrieval. The minimum relevance score takes effect only in semantic retrieval or result reranking; it is a filtering threshold and does not perform permission or entity isolation. A similarity setting cannot exclude unrelated business or unauthorized records.
Hypothetical paper exercise sample
The following is a paper exercise used to explain the checking method. It does not represent any real customer or actual test.
Material text one (excerpt from an advertising service contract): "Party A commissions Party B to place feed advertising from 1 March 2025 to 31 March 2025, with a daily budget not exceeding RMB 8,000; any excess requires written confirmation from Party A."
Material text two (excerpt from a settlement statement): "Settlement period 1 March 2025 to 31 March 2025, actual consumption RMB 236,000, of which single-day consumption on 18 March was RMB 11,200."
Draft sentence to be judged: "Party B completed the placement within the authorized scope, and consumption on 18 March did not exceed the agreed amount."
The fields required for judgment are the same subject (Party B), the same period (March 2025) and the same field (daily budget cap compared with actual single-day consumption). The contract states a daily budget not exceeding RMB 8,000 and requires written confirmation for any excess. The settlement statement states single-day consumption of RMB 11,200 on 18 March. The draft sentence claims no excess over the agreed amount, which is mutually exclusive with the cap value stated in the contract, and the materials do not state written confirmation from Party A. The draft sentence is unsupported. To determine whether written confirmation exists, a supplementary material search must be conducted within an authorized retrieval scope; when absent, the evidence gap is recorded.
Configuration and implementation checks
The table below lists only documented settings or implementation check items relevant to document parsing and chunking and contract and transaction material comparison on this page. The values are starting points for this hypothetical exercise and are not official default values.
| Configuration or check item | Starting point for this exercise and reason | Comparison method |
|---|---|---|
| Search mode | Hybrid retrieval is selected because the sample contains precise fields such as dates and amounts, and also natural language expressions | Observe whether numbering and amount questions hit the corresponding chunks |
| Citation limit | Set to a token amount that covers the contract clause and settlement statement in the same passage; the starting point is determined by sample passage length | Compare whether a single citation contains both the authorization period and the consumption amount |
| Minimum relevance | A lower starting value avoids filtering out chunks containing precise fields | Compare whether recalled content before and after adjustment includes the sample passages |
| Question optimization | The sample question has a complete subject, so it is not enabled, reducing one model call | Compare whether the retrieval terms change when enabled and when not enabled |
| Result reranking | Used when a rerank model is available and enabled by configuration, filtering by rerank score | Compare candidate ordering and scores before and after reranking |
| Chunking and index check | Check whether the contract clause and settlement statement each retain a complete judgment unit | Inspect whether the same chunk contains the condition and the content it governs |
Three common errors
- Treating a successful upload as retrieval already updated. After material changes, the actual training and index status must be checked; the upload action itself does not indicate that new chunks are retrievable.
- Treating a similarity setting as permission or entity filtering. The minimum relevance is a filtering threshold; retrieval does not perform authorization judgment, and unauthorized records cannot be excluded by lowering similarity.
- Treating the fused result after hybrid retrieval as a single vector ranking. RRF is a fusion process for mixed results, and the final cited content may come from multiple recall paths; ordering cannot be inferred from a single vector score.
How to verify the setup
- Input sample material text one and text two, and compare whether the subject, period and field in the draft sentence match the source text. When the draft sentence claims consumption on 18 March did not exceed the agreed amount, while text one states a daily budget not exceeding RMB 8,000 and text two states that day's consumption was RMB 11,200, the judgment is unsupported, and the condition is met.
- Input the retrieval term "March 2025 daily budget consumption" and compare whether the returned chunk contains both the authorization period and the amount field. When the returned content can be traced back to the corresponding passages of text one and text two, chunking meets the comparison need; if only the amount is returned without the period, record a chunk separation problem.
- Input an empty-result scenario and check the downstream design with an explicit paper input. When retrieval returns empty, record that state and check the downstream citation flow, without inferring that a fictional rule necessarily returns an empty result.