What the data for this category looks like
The core data for this category comes from advertising platforms, user profile databases, budget scheduling tables, and attribution tools. Daily documents include delivery plans, targeting setting rules, budget allocation plans, and conversion tracking configuration items. Most data updates happen via daily synchronization or when a single campaign launches. Document structures mostly use structured field combinations, with clear fields such as target audience tags, budget amounts, conversion event IDs, and delivery time periods. Unstructured content makes up a very small share of total content.
Constraints These Characteristics Impose on Vector Models and Indexes
The structured data of this category requires vector indexes to retain semantic associations between fields. Do not arbitrarily split associated fields such as targeting tags and budget values. High-frequency data updates require indexes to support incremental synchronization. This prevents performance losses from full index reconstruction. Clear field units and business meanings require retaining field context during vector generation. This stops numerical values with different units from being mismatched incorrectly. Adjust the maxToken parameter to control input length. This prevents long budget scheduling plans from being truncated. Individual data entries vary widely in length. Some targeting rules are only tens of characters long. Some budget scheduling plans span thousands of characters. The vector processing logic must support multi-length text.
How to Configure Settings
The following configuration items are written for the vector model and index module in FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Delivery plan documents are mostly structured field combinations. This length retains the complete semantics of a single business rule, and avoids splitting that breaks associated logic |
| Overlap Length | 100–200 characters | Adds contextual information across chunks. This ensures semantic coherence for cross-chunk content such as targeting settings and budget schedules |
maxToken | 8192 | Adapts to the maximum input limit of most open-source embedding models, and avoids text truncation that loses core business information |
defaultToken | Calibrated based on actual testing | Fills remaining tokens for individual data entries, ensuring complete vectors are generated for input text of varying lengths |
| Similarity Threshold | 0.75–0.85 | Matching for delivery plans requires precise alignment with targeting rules and budget parameters. This range filters low-correlation historical configuration plans |
| Single Recall Limit | Top 10 entries | Configuration references for advertising delivery plans usually do not require excessive historical data. Controlling the recall volume improves query efficiency |
Three Common Configuration Mistakes
- Issue: After configuring a vector model with the same name, the original configuration is overwritten, and multiple embedding models of the same type cannot be called simultaneously. Cause: Multi-model group configuration is not enabled. The system uses the model name as the unique identifier by default. Submitting duplicate named configurations directly overwrites existing entries.
- Issue: After structured data is chunked, targeting tags and budget values are split into different chunks. Associated logic is lost during matching. Cause: Chunk length is set too small. The semantic binding relationship of structured fields is not considered. Forced splitting breaks business context.
- Issue: Truncation occurs when generating vectors for long-text budget scheduling plans. Complete configuration information cannot be matched. Cause: The
maxTokenparameter is not adjusted to the maximum input length supported by the current model. Text exceeding the limit is directly truncated, losing core configuration information.
How to Verify Successful Configuration
- Upload a single delivery plan document, and check the chunk preview interface. Confirm that structured fields are not split across different chunks.
- Call the vector generation API, and check that the returned vector dimensions match the output dimensions of the selected embedding model.
- Run a test query, and verify that the number of recall results matches the set single recall limit.
- Submit a model configuration with the same name, and check if the system prompts duplicate configuration risks or supports group identifiers.
The values listed here are common starting points. They should be validated against local test samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-10-19.