What the documents look like
Inventory procurement and supplier records on a B2B marketplace draw on the platform itself, the national enterprise credit information publicity system, payment transaction records, and logistics tracking. The working documents are enterprise certification materials, inquiry and quotation sheets, electronic contracts, and platform settlement statements. Compared with wholesale market stall trading, this category carries more online certification and electronic contracts, and transactions do not depend on a physical stall. Records update daily and per order.
The four document types have different structures, so ingestion segmentation follows each structure. A condition and the content it constrains stay in the same segment. One inquiry and quotation sheet may map to several indexes, and retrieval merges them by original data. A successful upload is not a completed training run, and it is not an updated retrieval index.
Constraints at the document chunking step
Parsing and chunking distinguish original data from retrieval indexes. In FastGPT, the MongoDB dataset.datas table stores vector original data, and the indexes field records the corresponding vector IDs, so one set of data can correspond to multiple vectors. PostgresSQL stores vectors in a vector field with an HNSW index. Retrieval recalls vectors first, then uses vector IDs to find original data in MongoDB; entries matching the same original data are merged, and the vector score takes the highest value.
Segmentation completeness directly affects retrieval. When a condition, an explanatory note, and the content it constrains are split apart, the single segment is semantically incomplete and vector precision drops. In enterprise certification materials, the entity name, certification status, and validity period stay in one segment. In an inquiry and quotation sheet, the material, quantity, and quotation validity period stay in one segment. In an electronic contract, the signing parties, signing status, and effectiveness conditions stay in one segment. In a platform settlement statement, the settlement cycle, amount, and status stay in one segment.
Hypothetical paper exercise sample: the following paper materials are for checking segmentation and retrieval and involve no real customer or real company.
Source text one (enterprise certification materials): "The enterprise certification materials submitted by Supplier A state: the entity name is A Supply Chain Co., Ltd.; the certification status is pending review; the submission date is 10 March 2025; the certification validity period is not stated."
Source text two (inquiry and quotation sheet): "On 12 March 2025, A Supply Chain Co., Ltd. sent an inquiry and quotation sheet to Buyer B; the material is Type A bearings; the quantity is 500 units; the quotation validity period runs to 31 March 2025; the supply period is not stated."
Draft sentence to be judged: "A Supply Chain Co., Ltd. has passed enterprise certification, and a quotation validity period running to 31 March 2025 means the supply period runs to that date."
Judgment points: source text one states the certification status as pending review, while the draft sentence says enterprise certification has passed; same entity, same field, clearly mutually exclusive values, so this is a contradiction. Source text two does not state the supply period, and the draft sentence equates the quotation validity period with the supply period, which is unsupported. Pending review does not show an absence of operating authority, and a quotation validity period does not indicate a supply period.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | Reason and comparison method |
|---|---|---|
| Search mode | Hybrid retrieval | Enterprise certification materials contain exact terms such as entity names, and inquiry and quotation sheets contain natural language descriptions; hybrid retrieval uses semantic recall and full-text recall together and merges them with RRF. Compare recalled content against semantic retrieval alone. |
| Minimum relevance | 0.3 | A hypothetical starting value for this exercise, not an official default. It takes effect in semantic retrieval or with result reranking. Compare with 0.5 and observe whether recalled content contains unrelated fragments. |
| Reference limit | 2000 tokens | A hypothetical starting value for this exercise, not an official default. It controls the volume of quoted content by tokens. Compare with 3000 and check whether quoted content exceeds what is needed. |
| Question optimization | Enabled | In multi-turn dialogue, coreference resolution and question expansion can improve retrieval, at the cost of one extra model call and added latency. Compare recalled content in a follow-up question against the disabled setting. |
| Result reranking | Enabled | The rerank model performs a second ordering of candidate text with a 0-1 score that supports filtering. Compare text relevance against the disabled setting. |
| Segmentation method | Segmentation by document structure | The four document types have different structures, and structure-based segmentation preserves the completeness of conditions and the content they constrain. Compare segment semantic completeness against fixed-length segmentation. |
Three common errors
- Treating a successful upload as an updated retrieval index. After source text one is submitted, without confirming training and index status, retrieval may still return old content. Check the actual training status and index status rather than the upload record alone.
- Treating pending review as passed. Source text one states the certification status as pending review, while the draft sentence says enterprise certification has passed. Same entity, same field, clearly mutually exclusive values, so this is a contradiction. Compare the status field in the source text against the wording of the draft sentence.
- Treating a quotation validity period as a supply period. Source text two states a quotation validity period running to 31 March 2025 and does not state a supply period. The draft sentence equates the two, which is unsupported. Compare whether the source text states a supply period field.
How to verify the setup
- Input source text one and the draft sentence "A Supply Chain Co., Ltd. has passed enterprise certification." Compare the certification status field in the source text against the draft sentence. The source text states pending review and the draft sentence says passed; same entity, same field, clearly mutually exclusive values, so the judgment is a contradiction. If the status field is absent from the source text, record the result as unsupported.
- Input source text two and the draft sentence "a quotation validity period running to 31 March 2025 means the supply period runs to that date." Compare whether the source text states a supply period field. The source text does not state a supply period, and the draft sentence equates the quotation validity period with the supply period, so the judgment is unsupported. If the source text states a supply period consistent with the quotation validity period, the judgment is consistent.
- Input source text one and run separate retrievals with "A Supply Chain Co., Ltd.", "pending review", and "10 March 2025". Observe whether returned content contains that entity, status, and date. If a retrieval result is empty, check whether the data has completed training, whether the minimum relevance is too high, and whether the reference limit is too low. An empty result means no recall under the current retrieval conditions and does not mean the material does not exist.