What the data for this category looks like
Data comes primarily from project operation logs, public area inspection reports, facility maintenance ledgers, tenant lease contracts, and property fee collection records. Update frequency varies by business node: inspection reports are updated weekly, maintenance ledgers are updated per equipment maintenance cycles, lease contracts and collection records are updated monthly. Single documents contain fields including area code, equipment number, inspection time, issue severity, rectification results, and more. Some fields have fixed units, such as lease area (square meters), rectification time spent (hours), and collection amount (yuan).
What constraints these characteristics impose on vector models and indexing
Mixed multi-type fields require vector models to support joint encoding of structured fields and unstructured text, to avoid losing numerical and temporal information with single-text encoding. Fluctuations in incremental data share caused by uneven update rhythms require index systems to support incremental construction and dynamic updates, to avoid resource consumption from full reconstruction. Single documents have wide variation in length: from hundreds of words in inspection records to thousands of words in maintenance plan documents. This requires configurable adaptive chunking rules to ensure semantic integrity after chunking. Some fields have fixed units, so numerical fields must be standardized in advance to ensure consistency in the vector space.
How to Configure Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
chunk_size | 800–1200 characters | Adapts to the wide length variation of property documents, balances semantic integrity and vector encoding efficiency |
overlap_ratio | 10%–15% | Prevents semantic breakage after chunking long documents, adapts to business logic associations across chunks |
vector_db_type | pgvector / Qdrant / Milvus | Supports structured field filtering and incremental updates, adapts to the uneven update rhythm of property data |
embedding_model | bge-large-zh-v1.5 or a localized Chinese embedding model | Adapts to property-specific business terminology, improves the accuracy of vector encoding |
recall_top_k | Top 8–12 results | Covers associated investment research information across multiple areas and devices, balances recall coverage and retrieval efficiency |
similarity_threshold | 0.72–0.80 | Distinguishes similarity between similar business cases, avoids mistakenly recalling operation records from unrelated areas |
The parameter values provided on this page are all common recommended 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 settings.
Three Common Misconfigurations
- Symptom: Knowledge base indexing tasks remain in a running state with no progress updates. Cause: No incremental indexing rules are configured, and full indexing traverses a large volume of historical operation data, leading to task blocking.
- Symptom: Embedding model interface calls return a 503 status code, and logs show no available models in the
defaultgroup. Cause: No dedicated group is configured fortext-embeddingtype models, or there are insufficient model instances in the group to handle concurrent requests. - Symptom: A large number of inspection records from unrelated areas appear in retrieval results, with insufficient recall precision. Cause: No structured field filtering rules are configured, and no vector encoding association is made for fields such as area code and equipment number.
How to Verify Correct Configuration
- Run a chunking test for a single property document, check the semantic integrity of the chunked content, and adjust the corresponding configuration items to values that meet business requirements.
- Initiate a batch incremental indexing task, verify that the indexing task can be completed within the preset time, with no persistent blocking.
- Call the retrieval interface, pass the area code as a filter condition, confirm that only relevant documents from the corresponding area are returned, and verify that the structured filtering rules take effect.
- Check the embedding model call logs, confirm that no 503 status codes are returned, and that the model group configuration meets call requirements.
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.