What the Data for This Category Looks Like
The data for meeting minutes comes from audio-to-text transcripts of government and enterprise meetings, digitized paper records compiled by attendees, and standard documents exported from meeting management systems. Update cycles include generating a single set of minutes after a single meeting, or batch importing historical meeting minutes on a weekly or monthly basis. The document structure includes fixed fields such as meeting topic, start time, attendee list, discussion topics, resolution items, follow-up tasks, and more. The length of individual documents varies widely. All fields are structured or semi-structured text, with a single set of minutes as the basic processing unit.
What Constraints These Characteristics Impose on Embedding Models and Indexing
Multi-source input for meeting minutes creates differences in text formats. Audio-to-text transcripts contain colloquial redundant expressions. Structured exported content has standardized fields. This requires the embedding models and indexing link to support mixed-format text input. There are clear differences in fields across individual documents. Short fields such as attendees and resolution items are usually within a few hundred characters. Long fields such as discussion topics can be thousands of characters long. This requires matching differentiated chunking logic for different fields. Batch-imported historical minutes and newly generated real-time minutes place different requirements on the index's real-time update and batch construction capabilities. The system must adapt the indexing rhythm for both update scenarios.
How to Set the Configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Matches the average text length of discussion topics and resolution items in meeting minutes, avoids overloading single chunks |
| Chunk Overlap Length | 100–200 characters | Retains contextual connections across chunks, prevents the logic of meeting topics from being broken by chunking |
maxToken | 8192 | Adapts to the maximum input token limit of mainstream embedding models, supports long text chunks from meeting minutes |
defaultToken | 1536 | Matches the output dimension standard of most embedding models, ensures consistency of vector indexes |
| Similarity Threshold | 0.72–0.85, calibrated via actual testing | Meets the semantic matching accuracy requirements for resolution items and follow-up tasks in meeting minutes |
| Single Search Limit | Top 10 results | Covers the typical number of topics and resolutions in meeting minutes, avoids redundant recall |
Three Common Misconfigurations
- The symptom: After calling the pushdata API to upload chunk-mode data, the interface shows the indexing status for an extended period. The cause: Batch indexing trigger rules are not configured for the meeting minutes batch import scenario, leading to repeated index construction triggered by individual chunks.
- The symptom: Embedding model calls return an error indicating that the token does not have permission to use the specified model. The cause: The access key for the corresponding model is not correctly bound in the FastGPT configuration, or the key has not been granted permission for the specified embedding model.
- The symptom: Vector recall results do not match the actual topics of the meeting minutes. The cause: The chunk length is set too large, causing a single chunk to contain multiple unrelated meeting discussion contents, leading to semantic vector confusion.
How to Confirm Proper Configuration
- Upload a standard-format meeting minutes test sample, verify that the vector index construction progress is completed within 5 minutes in FastGPT v4.17.0, with no prolonged hanging status.
- Call the embedding model test interface to confirm that the returned vector dimension matches the configured
defaultTokenparameter. - Enter a query related to the meeting minutes, check that the similarity scores of the recall results fall within the configured threshold range.
- Batch import 10 sets of historical meeting minutes, confirm that the index construction tasks are completed according to the batch rules, with no duplicate index entries.
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 2026-09-28.