What the data for this category looks like
Thermal marketing data mainly comes from offline thermal spot passenger flow collection systems, user online behavior tracking APIs, and store operation backends. Data updates occur hourly or in real time. Some core spots support minute-level refresh. Each data entry includes fields such as unique spot identifier, spot name, collection period, passenger count, average stay duration, geographic coordinates, and user interaction tags. Passenger count is measured in passenger trips. Stay duration is measured in seconds. Geographic coordinates use decimal format. Collection time uses the ISO 8601 standard format.
What constraints these characteristics impose on model integration and configuration
The hourly or real-time update rhythm of thermal marketing data requires configuring an incremental index update strategy aligned with the data refresh cycle during model integration. This prevents recalling expired data. Multi-field structured features including geographic coordinates and interaction tags require configuring field mapping rules. Core fields such as spot identifiers and behavior tags must be prioritized as model inputs. Standardized units for passenger count and stay duration require unifying data formats during preprocessing. This avoids model inference errors caused by unit mismatches. The unique spot identifier field must be configured as the knowledge base primary key. This prevents duplicate indexing of data for the same spot.
How to set the configuration
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
embedding_batch_size | 32–64 entries | Single thermal data entry contains multiple fields. A batch that is too large triggers memory limits, while a batch that is too small reduces index construction efficiency |
index_update_interval | 1 hour | Matches the hourly update rhythm of thermal data, balancing index latency and server resource usage |
max_context_length | 800–1200 characters | The total field length of a single thermal data entry is moderate. An overly long context window introduces redundant fields that interfere with model inference |
unique_key_field | point_id | The unique spot identifier is the core identification field for thermal data, preventing duplicate indexing of historical data for the same spot |
embedding_model_type | locally deployed M3E series | Structured thermal data requires a lightweight model with accurate semantic understanding. Local deployment avoids network latency from cloud calls |
parse_field_separator | , | Thermal data mostly uses comma-separated structured formats, with better compatibility with default parsing rules |
The parameter values provided on this page are 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.
Three common errors
- When calling a locally deployed M3E model, a connection timeout error occurs. Investigation shows the
embedding_api_addressconfiguration is missing the http:// protocol prefix. After version 4.8.19, the local model address configuration item added protocol verification rules. Incorrect configuration causes connection failures. - When configuring oneapi as an embedding model, the embedding interface used by the index model cannot be specified separately. This happens because the "Independent Index Model Configuration" switch is not enabled. The system defaults to reusing the chat model's configuration parameters, overwriting the oneapi embedding interface settings.
- Duplicate spot recall results appear when retrieving thermal marketing data, and logs indicate the primary key field is not matched correctly. This occurs because
unique_key_fieldis not configured aspoint_id. The system uses the file path as the primary key by default, causing multi-period data for the same spot to be recognized repeatedly.
How to confirm the configuration is complete
- View the knowledge base index logs to confirm all fields of each thermal data entry are parsed correctly, with no missing or abnormally formatted fields.
- Initiate an embedding test for a single data entry, and verify that the vector dimension of the embedding result matches the standard dimension of the selected model.
- Manually trigger an incremental index, and confirm the index update execution time matches the configured
index_update_interval. - Retrieve historical data for a specified spot, and confirm there are no duplicate entries in the recall results, and that field units conform to preset standards.
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.