What this category’s data looks like
Internal policy data comes from officially released internal enterprise policy documents, and covers formats including PDF, Word, and online collaborative documents. Update schedules are irregular, and are triggered when new policies are released, old policies are revised, or old policies are revoked. Document structures typically include fields such as section titles, clause numbers, effective dates, scope of application, and responsible parties. The length of single documents varies widely, ranging from a few pages to dozens of pages. Content is mostly structured text, with some accompanying charts or attachments.
What constraints these characteristics impose on vector models and indexing
The multi-format source of internal policies requires the indexing preprocessing step to support parsing common document formats such as PDF and Word, and to be compatible with some attached chart content. The irregular update schedule requires the indexing system to support incremental updates, to avoid resource waste caused by full rebuilds. Structured fields such as effective dates and scope of application in documents can be combined with vector retrieval to enable precise filtering and narrow the recall range. The wide variation in single document length requires that the chunking strategy not rely solely on fixed length, and must balance clause integrity and contextual coherence.
How to configure
| Configuration Item | Recommended Approach | Rationale |
|---|---|---|
embedding_model | Prefer text-embedding-v3; multimodal-embedding-v1 is an option for scenarios with many charts | Internal policies primarily consist of text content, so general-purpose text embedding models have better adaptability. Multimodal models can accommodate policy attachments with charts |
chunk_size | 800–1200 characters | The length of individual clauses in internal policies mostly falls between 500 and 1000 characters. Chunk length covers individual clauses while retaining contextual connections |
chunk_overlap | 100–150 characters | Avoid truncating clause context across chunks, and ensure semantic coherence during retrieval |
recall_top_k | Top 5–8 results | Internal policy compliance question answering requires precise clause matching. Too many recall results increase subsequent filtering costs |
similarity_threshold | 0.75–0.85 | Filter low-match irrelevant policy content, and ensure the accuracy of compliance question answering |
index_refresh_interval | Calibrated based on actual testing | Adapt to the irregular update schedule of internal policies, and adjust the incremental index trigger timing as needed |
The parameter values provided on this page are general recommendations used as a starting point for configuration. Actual values are affected by material form, data volume, and business rules. Specific issues require individual analysis. It is recommended to test on your own samples before finalizing values.
Three common mistakes
- Phenomenon: An
invalid model typeerror is prompted after saving the vector model configuration. Cause: The API key and model identifier of the corresponding service provider are not correctly bound in the FastGPT vector model channel configuration, or an unauthorized vector model type is selected. - Phenomenon: The number of recall results after index construction is far lower than expected. Cause:
chunk_sizeis set too small, causing long clauses in internal policies to be excessively truncated, breaking semantic connections and preventing correct recall. - Phenomenon: Manual index update cannot be triggered after deployment via docker-compose. Cause: The
auto_index_triggerparameter is not enabled in the vector service configuration ofdocker-compose.yml, or the correct index storage directory is not mounted, preventing index files from being written.
How to confirm the configuration is complete
- Upload an internal policy document, check the parsed chunking results, and confirm that the chunk length falls within the preset
chunk_sizerange. - Launch a compliance question answering test, enter a question that highly matches a policy clause, and check whether the similarity scores of the recall results fall within the interval set by
similarity_threshold. - Modify an uploaded policy document, trigger reindexing, confirm that a corresponding update task appears in the index queue, and that the new content can be correctly recalled after completion.
- Check the vector service runtime logs, confirm that there are no error messages such as
embedding failedorindex build failed, and that the model call success rate is at a normal level.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-14.