What the data for this category looks like
Property management marketing content targeting the financial industry mainly comes from investment promotion brochures, owner service notices, quarterly event notifications, property fee standard documents, park service rules, and similar materials from various property projects under financial institutions. Update frequency is irregular, triggered by new property launches, fee standard adjustments, or offline event holdings. Each document has a relatively fixed structure, including fields such as project name, regional location, service content, fee standard, event time, and contact information. The unit for fee standards is mostly yuan/square meter·month, area unit is square meter, and time uses the YYYY-MM-DD format.
Constraints on knowledge base retrieval and recall from these characteristics
Data sources are scattered across property projects of different financial institutions. Independent knowledge bases must be split by project to avoid cross-institution content confusion. Update frequency is irregular, so incremental update trigger rules need to be configured to ensure new content is included in the retrieval scope in a timely manner. Documents contain clear numerical and time fields, so structured field retrieval and time range recall must be supported to improve matching accuracy. Each document is moderately long but includes a large number of similar business expressions, such as property fee descriptions for different projects. Retrieval matching rules must be adjusted to distinguish semantic differences. Retrieval filtering requirements that comply with financial industry content regulations must also be adapted.
Configuration Settings
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
recall count | Top 8–12 entries | Property management marketing documents have short individual content but large volume. Too many recalled entries will exceed the context window, while too few will miss relevant information |
similarity threshold | 0.72–0.80 | Marketing content contains a large number of similar expressions (such as property fee descriptions for different projects). A threshold that is too low will introduce irrelevant results, while a threshold that is too high will miss accurate content with slightly lower matching degrees |
chunk length | 800–1000 characters | Individual marketing documents (such as event notifications) are moderately long. Chunks that are too long will lose local semantic connections, while chunks that are too short will destroy complete business logic |
PARSE_FILE_TIMEOUT_SECONDS | 120 seconds | Some uploaded PDF park planning documents have a large number of pages, resulting in long parsing time. Insufficient default duration will cause parsing failures |
UPLOAD_FILE_MAX_SIZE | 500 MB | Supports uploading batch historical project marketing documents, avoiding inability to upload due to oversized single files |
reranked recall count | Top 4–6 entries | Perform secondary screening on recalled results to retain core content that best fits user queries, adapting to the precise matching needs of marketing scenarios |
The parameter values provided on this page are all conventional recommendations used as starting points for configuration. Actual values are affected by material form, data volume and business rules. Specific issues require specific analysis, and it is recommended to test on your own samples before finalizing.
Three common mistakes
- Symptom: Calling the recall test API returns {"code":514,"statusText":"invalid kb id"}, and the error persists after confirming the curl configuration is correct. Cause: For the v4.8.12-alpha open source version, the property management project's marketing knowledge base was not separately bound to the corresponding application, or the incoming kb_id belongs to a deleted knowledge base.
- Symptom: The retrieved marketing content does not match the user's query, and a large number of irrelevant park maintenance notifications appear. Cause: The similarity threshold was set too high, causing accurate content with slightly lower matching degrees but strong semantic association to be missed.
- Symptom: Uploaded docx format event notifications display empty fields after parsing. Cause: The FastGPT docx parsing plugin was not enabled, or the document contains encrypted content leading to parsing failure.
How to confirm the configuration is complete
- Enter the FastGPT knowledge base management page, select the corresponding property management marketing knowledge base, run a manual recall test, and verify whether the number of returned results matches the configured requirements.
- Upload a test property fee standard document, check whether the parsed fields include preset project name, fee standard, and other content, confirming that the parsing logic operates normally.
- Call the official recall test API, pass the correct kb_id and test query terms, and check whether the returned result has a status code of 200 with no abnormal error messages.
- Adjust the similarity threshold, run retrieval again, and observe whether the proportion of irrelevant content in the results matches the expected adjustment effect.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-14.