What data for this category includes
Data sources for image annotation projects include image samples for annotation, annotation data exported from annotation platforms, project-specific category dictionaries, quality inspection reports generated from spot checks, and rework lists. Data updates follow a daily schedule or the completion rhythm of each batch of annotation tasks. Daily produced documents include annotation specifications, bounding box sample files, quality inspection result reports, and rework task lists. Data fields include target bounding box coordinates, category labels, spot check markers, rework status, and others. Some fields use pixels and quantity as units.
Constraints imposed by these characteristics on multi-turn dialogue and prompt engineering
Diverse data sources and dynamic update requirements mean multi-turn dialogue must clearly separate visual annotation needs from text rules. Do not mix annotation rules with descriptions of visual examples. High-frequency data updates require support for synchronizing the latest category dictionaries or rework requirements during dialogue. Avoid using expired annotation standards. Multiple daily document types require prompts to specify the format for referencing specific documents. This ensures the annotation process follows unified specifications. The most time-consuming steps of bounding box calibration and occluded sample judgment require multi-turn dialogue to focus on visual details. Vague text descriptions must not interfere with annotation decisions.
Configuration settings for FastGPT v4.17.0
The following configuration items apply:
| Configuration Item | Recommended Approach | Rationale |
|---|---|---|
chatId | Bind to the exclusive ID of the annotation project | Ensure multi-turn dialogue contexts for the same project are not confused, and avoid cross-project data mixing |
messages | Split dialogue turns by annotation task batches, retain only the current batch's annotation rules and examples per turn | Avoid context overload, focus on the current batch's annotation requirements |
variables | Pass in the current batch's category dictionary, bounding box sample path, and quality inspection threshold | Dynamically load the latest project data, adapt to daily or per-batch update rhythms |
| Dialogue history retention turns | First 3 turns | Focus on the core rules of the current annotation task, avoid interference from irrelevant historical data |
| Maximum context length | 800–1200 characters | Adapt to the conventional length of annotation documents, avoid exceeding the model's processing limit |
Three common configuration errors
- Phenomenon: The API call returns a log showing the token field is "fastgpt", and the model cannot be called normally. Cause: The API call token parameter was not configured correctly, causing the token to be incorrectly replaced with the default value.
- Phenomenon: Redundant historical annotation data appears in multi-turn dialogue, leading to model output that does not meet the current batch's annotation requirements. Cause: The retention range of
messageswas not restricted, and excessive expired annotation task data was loaded. - Phenomenon: Context management logic differs across nodes in the workflow, causing deviations in annotation results for the same project. Cause: The rule for binding
chatIdwas not unified, and different nodes used different context management parameters.
How to confirm configuration is correct
- Call the API and check the returned log to confirm
chatIdmatches the exclusive ID of the current annotation project, with no cross-project data mixing. - View the dialogue history to confirm only the current batch's annotation rules and examples are retained, with no redundant historical data.
- After passing in the
variablesparameter, verify whether the model can correctly call the latest category dictionary and bounding box samples. - Test multi-turn dialogue to confirm each call's context retention rules comply with the configured requirements.
The values provided 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.