What the data for this category looks like
This category’s core data comes from internal enterprise procurement systems, supplier quotation documents, budget sheets, and contract systems. Daily documents include purchase requisitions, price comparison records, formal procurement contracts, and acceptance documents. Data updates follow individual procurement workflows. Related documents for a single purchase are updated after workflow nodes are completed. Bulk cross-month procurement data is synchronized once per month. Document fields include purchase order numbers, supplier names, budget amounts, contract clause numbers, and other relevant fields. Units include Chinese Yuan, procurement cycle days, device models, and other standard units.
What These Characteristics Impose on the Knowledge Base Retrieval and Recall Link
Data sources are scattered across multiple internal systems. Retrieval configurations must support cross-dataset grouping association to avoid retrieving irrelevant non-procurement documents. Update frequency follows individual procurement workflows or monthly intervals. High-frequency real-time incremental indexing is not required. A periodic full synchronization indexing strategy can be used. Document structure varies widely. Some documents are short purchase request texts, while others are long contract clauses. Matching logic must adapt to different text lengths. Fields include structured content such as amounts and identification numbers. Retrieval must prioritize precise matching of business-relevant fields to prevent irrelevant results from interfering with pre-sales and technical support judgments.
How to Configure Settings
These configurations apply to FastGPT v4.17.0 and later versions.
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the field density and text length of IT procurement documents, avoids retrieving irrelevant price comparison records or contract clauses |
limit | Top 8 entries | Covers multiple procurement documents that may be involved in a single pre-sales Q&A, while controlling the total context length |
searchMode | Full-text retrieval | Adapts to mixed text structures such as purchase requests and contracts, supports precise cross-field matching |
usingReRank | Enabled | Optimizes ranking of long document retrievals, meets business needs for accurate matching of contract clause verification |
datasetSearchUsingExtensionQuery | Enabled | Optimizes reference resolution, resolves ambiguous issues such as "this quotation" or "that contract" in pre-sales Q&A |
collectionIds | Configured by business grouping | Distinguishes datasets such as purchase request libraries and supplier quotation libraries, avoids interference from cross-category retrievals |
Three Common Mistakes in Configuration
- Phenomenon: The knowledge base search node returns empty results or no matching content. Cause: No valid
collectionIdsspecified to associate with the corresponding business dataset, or retrieval-related parameters are not configured correctly. Failure to configureAIPROXY_API_ENDPOINT(root address of the AI Proxy service, do not append/v1or other path suffixes) andAIPROXY_API_TOKEN(AI Proxy administrator token, consistent with server configuration and non-empty) on the server will cause startup verification failure, and the service cannot start normally. - Phenomenon: Retrieval parameters are automatically restored to default values after saving, and custom configurations cannot be retained. Cause: Configurations are not correctly stored on the server, or the workflow configuration is not bound to the access permissions of the corresponding dataset.
- Phenomenon: After uploading a PDF document, only text content is retrieved, and parsed text from images is not included. Cause: The OCR parsing configuration for the document is not enabled, or the extended retrieval text extraction logic is not turned on.
How to Confirm Configurations Are Correctly Set
- Check the server-side deployed environment variables, confirm that
AIPROXY_API_ENDPOINT(root address of the AI Proxy service, do not append/v1or other path suffixes) andAIPROXY_API_TOKEN(AI Proxy administrator token, consistent with server configuration and non-empty) are configured. - Send a test retrieval request with ambiguous reference questions, verify that corresponding procurement documents can be correctly associated, and resolve reference resolution issues.
- View the number of retrieval return results and similarity scores, confirm that they fall within the configured
limitandsimilarityvalue ranges. - Upload a test PDF document, verify that both text within the document and parsed content from images can be retrieved.
The values provided are common starting points and should be evaluated against the reader’s own sample data.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-10-19.