What the data for this category looks like
Business process data for telecommunications engineering construction primarily comes from documents such as project approval documents, construction drawings, acceptance specifications, and operation manuals. Update frequency adjusts based on project progress or industry standard revisions. Single project documents include structured fields like project number, construction location, construction unit, and approval document number, as well as unstructured content such as construction process descriptions, equipment parameters, and engineering drawings. Field units use professional engineering units including kilometers, square meters, days, and ten thousand yuan. Individual document lengths range from hundreds to thousands of characters.
What constraints these characteristics impose on knowledge base retrieval and recall
The presence of structured fields and professional units requires retrieval to match field semantics and unit consistency, to avoid invalid recall caused by unit mismatches. Unstructured engineering drawings and long text content increase vector embedding computation load, so the number of recall results must be controlled. Separate project documents require retrieval to limit the scope to the corresponding project’s dataset, to avoid interference from irrelevant cross-project content. Frequent use of abbreviations and professional terms requires retrieval to support expanded queries, covering complete professional expressions. Image document links must be fully retained to avoid truncation during recall.
Configuration Settings
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72-0.85 | Business issues related to telecommunications engineering construction mostly involve precise approval document numbers and construction parameters. A threshold that is too low introduces irrelevant content, while a threshold that is too high misses valid results. This range is calibrated based on actual testing. |
limit | Top 8 entries | Questions about business process guidance usually involve multi-step approval nodes. Around 8 recall results cover the main process nodes and avoid excessive redundancy. |
searchMode | vector | Documents for telecommunications engineering construction contain a large number of professional terms and structured fields. Vector retrieval can match semantically similar professional expressions, adapting to professional scenario needs. |
usingReRank | Enabled | Documents for telecommunications engineering construction contain a large amount of similar construction specification content. Re-ranking filters recall results with weak semantic relevance and improves accuracy. |
datasetSearchUsingExtensionQuery | Enabled | Business process questions often include professional abbreviations. Expanded queries supplement the full expressions of abbreviations, improving recall coverage. |
collectionIds | Fill in the dataset ID of the currently processed project | Business handling for telecommunications engineering construction is usually carried out on a single project basis. Limiting the dataset avoids interference from irrelevant cross-project content. |
Three Common Mistakes
- Phenomenon: The knowledge base returns incomplete image links, with some characters truncated. Cause: Image links in imported documents contain unescaped special characters. The knowledge base recall process does not perform full encoding on links, leading to truncation during output.
- Phenomenon: When testing a workflow, only the first question can associate knowledge base content, and subsequent similar questions have no knowledge base references. Cause: The workflow is not configured with session context passing. Subsequent requests do not carry the correct
collectionIdsparameter, resulting in an empty retrieval scope. - Phenomenon: After importing bilingual documents, the large language model does not call knowledge base content when encountering Chinese questions. Cause: Multilingual retrieval configuration is not enabled. The vector model only generates embeddings for single-language text, causing Chinese queries to fail to match Chinese bilingual document fragments.
How to Confirm the Configuration is Correct
- Upload a test document containing a project number and construction location, initiate a query that includes the number, and check if the corresponding document fragment is included in the recall results.
- Initiate a query that includes a professional abbreviation, such as "FTTH construction process", and check if content containing "fiber to the home" is recalled.
- Import an engineering drawing with a complete link, initiate a query that includes drawing keywords, and check if the returned image link is complete and not truncated.
- Initiate 3 consecutive business questions for the same project, and check that each request’s returned results include knowledge base content for the corresponding project.
All provided configuration values are common starting points. Adjust them based on your own specific samples and testing.
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.