What this category’s data consists of
Data for this category primarily comes from image samples, annotation platforms, category dictionaries, and spot check results. Daily associated documents include annotation specifications, bounding box examples, quality inspection reports, and rework lists. Data updates follow a daily or per-project batch schedule. Most documents focus on visual target position annotation, category definitions, and quality inspection results. Fields include annotation bounding box coordinates, category labels, rework markers, spot check confidence scores, and similar metadata. This setup requires handling annotation association information for single or batch images. It is distinct from annotation projects that only process text content.
Constraints on context and token processing
The data characteristics of this category impose multiple constraints on context and token processing steps. Batch image sample annotation metadata includes structured information such as bounding box coordinates and category labels. Token usage fluctuates significantly for single-batch data. Strict control of total context recall volume is required. The daily or per-batch update feature requires binding project batch identifiers during context association. This prevents cross-batch data confusion. Multiple types of associated documents, including annotation specifications and quality inspection reports, must have recalled content filtered by business scenario. This avoids irrelevant information occupying token quotas. Association relationships between visual annotation fields must also be preserved. This prevents loss of target position information after content splitting.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
maxContext | 8000–12000 token | Applies to FastGPT v4.17.0. Image annotation project metadata includes bounding box coordinates and batch image information. Token usage is higher than text annotation. Sufficient context space must be reserved. |
quoteMaxToken | 2000–3000 token | Effective content of single annotation-related documents (such as bounding box examples, quality inspection reports) should not be overly long. This avoids exceeding the token limit for single citations. |
| Single search upper limit | Top 3–5 entries | Associated documents for image annotation projects are mostly categorized by batch. Excessive recall leads to context redundancy and token quota consumption. |
| Context batch filtering | Match by project batch | Data for this category updates daily or per batch. Filtering non-current project context data reduces invalid token consumption. |
| Context recall similarity threshold | 0.75–0.85 | Semantic relevance of visually annotated keywords (such as target categories, bounding boxes) must fall within a reasonable range. This prevents recall of irrelevant annotation specifications or rework records. |
Three common configuration errors
- Scenario: A token limit exceeded error triggers during a single response even when
quoteMaxTokenis set to 2000 token. Cause: The structured token usage of image annotation metadata is not accounted for. Only the single citation upper limit is configured. The global context quota formaxContextis not adjusted synchronously. - Scenario: Context is lost during a conversation. Previous annotation rework records cannot be associated with the current thread. Cause: Context filtering tied to project batches is not enabled. Cross-batch context data is incorrectly recalled or overwritten.
- Scenario: After annotation specifications are imported into the knowledge base, the system automatically splits paragraphs containing bounding box coordinates. This causes loss of target position association information in subsequent recalls. Cause: Rules prohibiting splitting for structured visual annotation content are not configured. The default splitting logic breaks the association between bounding boxes and corresponding images.
How to verify correct configuration
- Import a single batch of annotation samples and associated documents. Trigger an annotation quality inspection report generation process. Check that context recall content only includes data relevant to the current batch.
- Review token consumption statistics in system logs. Confirm that total token usage for a single request does not exceed the preset
maxContextquota. - Test multiple consecutive conversation rounds. Verify that previous rework lists, bounding box examples, and other content are correctly associated with the current conversation context.
- Perform a knowledge base import operation on a paragraph containing bounding box coordinates. Confirm that the paragraph is not split, or that associated target position information is retained after splitting.
The values provided here are common starting points. They should be evaluated against the reader’s own sample data.
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.