What the data for this category looks like
Daily data for image annotation projects includes four types of documents: annotation specifications, bounding box examples, quality inspection reports, and rework lists. Data sources are image samples, category dictionaries, and spot check results from the annotation platform. Data is updated daily or per batch. Core fields include visual target position coordinates, category labels, annotator information, and review comments. Unlike text annotation projects, this category focuses on visual target positioning and classification, with no pure-text semantic annotation content.
Constraints these characteristics impose on context and token usage
Daily data for image annotation projects centers on visual target positioning and classification, including annotation specifications, bounding box examples, quality inspection reports, and rework lists. Bounding box examples and quality inspection reports often contain coordinate descriptions, category labels, and similar content. The token consumption per piece of content is higher than that of pure-text annotation data. This category of data is updated daily or per batch, and older context content expires alongside new batch annotation tasks. If reasonable context cleanup rules are not set, redundant historical records will quickly exhaust token quotas. Annotation reviews use associated documents from the current batch. Excessive recall of historical data wastes tokens and disrupts the accuracy of current reviews. Therefore, context recall scope and token usage must be constrained based on the characteristics of this category. Excessive recall of historical data should be avoided.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
maxContext | Last 3–5 annotation review conversations + current batch documents | Annotation reviews for image annotation projects only rely on associated content from the current batch. Excessive historical conversations occupy unnecessary tokens |
quoteMaxToken | 1200–1800 token | The text description length of a single annotation specification or bounding box example typically falls within this range, which can fully cover core content without exceeding model limits |
| Context auto-cleanup trigger condition | Triggered by annotation batch updates | Annotation data is updated daily or per batch. Cleaning by batch can accurately remove expired context content |
| Single review input token limit | 8000 token | Combined with total token consumption of current batch bounding box examples, quality inspection reports, and review conversations, this avoids model response timeouts |
| Context recall count | Last 2 most recent quality inspection reports | Image annotation reviews primarily reference recent batch quality inspection results. Excessive recall wastes token resources |
Three common misconfigurations
- Issue: After submitting an annotation review task, the system prompts token limit exceeded, and returned content is incomplete. Cause: The value of
maxContextis not restricted, and annotation documents from multiple batches are all included in the context, exceeding the token capacity limit of the model. - Issue: The input and output token usage of annotation review conversations cannot be viewed separately, only the total token value is displayed. Cause: Fine-grained token statistics configuration is not enabled in FastGPT v4.17.0, and only the basic token calculation interface is called.
- Issue: Old batch annotation review records can still be retrieved after triggering a context clear operation. Cause: The configured cleanup rule only triggers based on time, and does not combine with annotation batch update status, resulting in expired content not being properly removed.
How to verify correct configuration
- Submit an annotation review task that includes bounding box examples and quality inspection reports, view the system returned token usage details, and confirm that input and output tokens are displayed separately.
- Manually trigger a batch-based context cleanup operation, retrieve historical records, and confirm that old annotation review content beyond the current batch has been removed.
- Upload an annotation specification document that exceeds the recommended length of
quoteMaxToken, verify whether the system triggers automatic truncation or a token limit exceeded prompt. - Adjust the value of
maxContexttoLast 1 annotation review conversation, submit a test task, and confirm that the context only contains the specified number of historical contents.
The values provided are common starting points and should be validated against locally tested 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.