What the data for this category looks like
The data for this category primarily comes from advertising platforms, asset libraries, user interaction data, and conversion data. Daily processed documents include creative scripts, asset test sheets, delivery data reports, and iteration checklists. Data updates follow a daily or weekly cadence. Core processing stakeholders include in-feed ad optimizers, creative teams, and growth teams. The most time-consuming steps focus on creative variable design, asset fatigue assessment, and test conclusion summarization. Teams must compare the combined performance of images, videos, and copy. Errors may lead to reduced click-through rates, higher delivery costs, or asset policy violations.
What constraints do these characteristics impose on vector models and indexing workflows
Multi-source heterogeneous data sources require indexes to support cross-format text, metadata, and metric field association. Daily or weekly update cadences require indexes to adapt to incremental update and periodic full refresh mechanisms. The core need for creative combination comparison requires vector models to capture association features across multimodal content. The fast-paced iterative testing cycle requires index recall efficiency to match short-cycle retrieval response times. The risk of asset policy violations requires index-associated metadata fields to accurately align with compliance check rules, preventing non-compliant assets from being included in valid comparison scopes.
How to configure settings
The following configurations are based on the vector model and indexing logic of FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Matches the single-paragraph text length of creative scripts and asset test sheets, avoiding cross-topic chunking that disrupts creative logic |
| Overlap Length | 100–200 characters | Preserves contextual associations for creative combinations, preventing chunking from losing association information between variables |
| Similarity Threshold | 0.75–0.85 | Filters low-relevance asset test data, focusing on valid creative comparison results |
| Single Search Limit | Top 10 results | Adapts to the quick review needs of test conclusion summarization, avoiding redundant information interfering with decision-making |
defaultToken | 4096 | Matches the input length limit of mainstream vector models, adapting to long-text content in delivery data reports |
maxToken | 8192 | Supports integrating complete context across multiple iteration checklists and test data, meeting cross-document association analysis needs |
Three common configuration mistakes
- Symptom: Search result sorting does not match expectations, and results are not displayed in similarity priority order. Cause: No similarity threshold configured to match category requirements, or no association with creative combination metadata fields for secondary sorting.
- Symptom: An error is triggered when selecting a vector model other than ada-002 when creating a knowledge base. Cause: No access key and interface address for the corresponding vector model configured on the platform, or the selected model does not adapt to the text format requirements of the current knowledge base.
- Symptom: Auxiliary data is not correctly vectorized during incremental index updates. Cause: The index switch for auxiliary data is not enabled, or auxiliary data fields are not included in the vector extraction scope.
How to confirm correct configuration
- Submit a single creative script test data entry, view vector extraction logs, and confirm that chunk length and overlap length match preset configurations.
- Initiate a retrieval request for an asset combination, verify that the return result sorting logic aligns with the configured similarity threshold rules.
- Trigger an index update after importing auxiliary data, check that the index list includes the metadata fields of the auxiliary data.
- Switch the vector model and initiate a test retrieval, confirm no errors occur and the vector generation process operates normally.
The values provided here are common starting points, and should be measured against the reader's own 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.