What Data Looks Like for Text Annotation Projects
Core data for text annotation projects comes from customer corpora, content output by annotation platforms, specialized vocabularies, and quality inspection sampling results. Daily maintained documents include annotation specifications, sample packages, quality inspection reports, and problem feedback forms. Data updates occur daily or synchronized per batch annotation task. Each document has clear fields:
- Annotation specifications include tag definitions and boundary descriptions
- Sample packages contain text to be annotated and corresponding tag fields
- Quality inspection reports record sampling deviations and rectification items
- Problem feedback forms record annotation abnormalities and customer requests
Constraints for Knowledge Base Retrieval and Recall
Scattered data sources and frequent updates require the retrieval link to support real-time synchronous recall of multi-source documents. This avoids discrepancies between annotation specifications and actual annotation requirements caused by data lag. Diverse document types with clear fields require retrieval to accurately match semantic content such as tag definitions and sample boundaries. Data updated in batches or daily requires the retrieval link to support incremental knowledge base synchronization. Content from quality inspection reports and problem feedback forms must be quickly recalled for real-time reference during difficult sample arbitration and consistency sampling inspection. Incomplete retrieval results can affect annotation quality and delivery progress.
Configuration Settings
This configuration guide is based on FastGPT v4.17.0.
| UI Setting Item | Recommended Value | Rationale |
|---|---|---|
| Similarity Threshold | 0.75–0.85 | Matches tag definitions and semantic boundaries in annotation specifications. Avoids introducing irrelevant results if set too low, and misses precise matching professional content if set too high. |
| Single Search Limit | 6–10 items | Covers common question types for single batch annotation tasks, while controlling context length to adapt to model input restrictions. |
| Retrieval Mode | Semantic Retrieval | Adapts to semantic query needs for text annotation, such as tag boundary explanation and sample intent recognition scenarios. |
| Reranking Switch | Enabled | Performs secondary sorting on recall results to improve the display priority of documents related to difficult samples. |
| Query Optimization Switch | Disabled | Avoids additional rewriting of the original text of annotation samples, preserving the original format of annotation data. |
| Retrieval Scope | Only current project-specific knowledge base | Restricts retrieval to documents within the current project’s annotation specifications, sample packages and other materials, avoiding interference from cross-project documents on retrieval results. |
| Document Upload Restriction | Only text documents | Aligns with data source requirements for text annotation projects, avoiding non-text content disrupting the retrieval process. |
Three Common Misconfigurations
- The interface prompts
invalid configuration parameter name "hnsw.iterative_scan". This occurs when a retrieval configuration item outside the FastGPT official parameter list is mistakenly used, and compliant specified parameters are not adopted. - The number of knowledge base search results does not match the set single search limit. This is caused by failing to correctly restrict the retrieval scope, leading to retrieval covering non-current project knowledge base documents and returning extra irrelevant results.
- Extra spaces and capital letter modifications appear in model outputs. This is caused by enabling the query optimization switch, which unnecessarily rewrites the original text of annotation samples, violating the original format requirements for annotation data.
How to Confirm Proper Configuration
- Upload the current project’s annotation specifications, sample packages, quality inspection reports and problem feedback forms to create a dedicated knowledge base, and confirm that the retrieval scope is bound to this knowledge base.
- Initiate a query for tag boundary explanation, and check whether the matching degree of returned results falls within the preset range.
- View the number of retrieval results, confirm that it matches the set single search limit, and no extra irrelevant documents are returned.
- Test input a query containing sample questions, and check whether the model output retains the original text format, with no extra spaces or capital letter modifications.
- Check the upload path of knowledge base files on the Ubuntu system, and confirm that documents have been successfully synchronized to the platform.
The configuration values listed here are common starting points. Adjust them based on your own specific samples and use cases.
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.