What this category of data looks like
Business data for infrastructure construction originates primarily from construction logs, bidding documents, site disclosure documents, material submission forms, acceptance specifications, and similar materials. Data updates align with project phases: on-site data is synced weekly during active project phases, and all documents in the current cycle are archived after phase acceptance. Single documents include dedicated fields such as section number, construction date, acceptance standards, and material model. Units follow standard engineering conventions, including meters, cubic meters, tons, workdays, and others. Some documents contain scanned construction drawings and structured tables.
What constraints do these characteristics impose on the knowledge base retrieval and recall link?
Dedicated fields and units require precise matching of project dimensions during retrieval. Otherwise, the system retrieves content from irrelevant sections, reducing the accuracy of business processing. Document lengths vary widely, ranging from a few pages of site disclosures to dozens of pages of acceptance reports. When splitting documents, retain associations between fields such as section number and construction date to avoid breaking business logic. Frequently updated on-site data requires regular re-embedding to prevent old data from falling out of sync with current project progress. Scanned drawings and tabular documents must also be included in the retrieval scope, which increases preprocessing complexity.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.80 | Infrastructure construction documents contain a large number of technical terms and project-specific numbers. A threshold that is too low introduces irrelevant section content, while a threshold that is too high misses relevant phase documents. Calibrated based on actual testing. |
limit | Top 8–12 results | Relevant documents for a single project’s business process usually focus on around 10 phases, avoiding exceeding the model’s context processing limit. |
usingReRank | Enabled | Semantic similarity differentiation for professional documents is low. Reranking improves the accuracy of the top 3 results, adapting to the precise retrieval needs of business processes. |
datasetSearchUsingExtensionQuery | Enabled | Questions about business processes often contain vague expressions. Extended queries cover synonyms and project abbreviations, improving recall coverage. |
collectionIds | Pass the knowledge base collection ID of the current project | Infrastructure projects divide knowledge bases by section, filtering documents from non-target projects to avoid retrieving interfering content. |
searchMode | vector | Prioritize semantic recall to match the contextual association of business processes, adapting to the semantic matching needs of professional documents. |
Three Common Misconfigurations
- Phenomenon: Tabular construction submission documents cannot be selected in the knowledge base management interface. Cause: The automatic parsing switch for tabular datasets is not enabled. The system skips structured non-text format documents by default.
- Phenomenon: Recall rate drops sharply for multilingual business documents after switching the embedding model. Cause: Batch re-embedding is not performed for all historical documents, causing semantic space deviation between old and new embedding vectors.
- Phenomenon: Invalid parameter prompt is returned when calling the API to create a text collection. Cause: The parameter corresponding to the paragraph-first splitting mode is not included in the request as required by FastGPT v4.17.0, causing the system to fail to recognize the splitting rules.
How to Verify Correct Configuration
- View the knowledge base dataset list to confirm that the tabular documents of the target section have been correctly loaded.
- Initiate a business process query to check whether all recalled results come from the current project’s knowledge base collection.
- Compare recall results before and after switching the embedding model to confirm that the semantic matching degree of multilingual documents meets expectations.
- Call the API to test text collection creation to confirm that no parameter error prompts are returned.
The values provided here are common starting points and should be measured against the reader's own samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-28.