What the data for this category looks like
Condiment marketing content is used for customer acquisition in finance, insurance, and wealth management scenarios. Sources include official brand product detail pages, e-commerce platform product copy, offline promotional poster text, live streaming sales scripts, dealer promotion scripts, and ingredient description documents for new product development. Update frequency fluctuates with marketing campaigns. Updates peak during new product launches and holiday promotions. Routine maintenance occurs monthly to quarterly. Document structure covers two categories: short text and long text. Short text includes short video scripts and social media promotion copy, with most entries under 50 characters. Long text includes product manuals and complete live streaming scripts. Core fields include product name, ingredient composition, packaging specifications (such as grams, milliliters, bottles, bags), applicable cooking scenarios, limited-time promotion information. Some documents include transcribed text from product photos.
What constraints these characteristics impose on vector models and indexing
The high proportion of short text requires vector models to support short text encoding, to prevent semantic dilution caused by overly large model context windows. Structured fields (such as specifications and promotion information) make up a clear share. Structured extraction of fields must be completed before indexing. Otherwise, marketing content for the same category with different specifications will have insufficient vector differentiation. Update frequency varies widely, with both batch updates and single-item additions. The indexing system must support incremental updates to avoid time delays from full reconstruction. Some documents include transcribed ingredient list text from images, which contains unstructured redundant content. Invalid noise must be filtered during the parsing stage, otherwise semantic accuracy during vector generation will be compromised.
How to set the configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Chunk size | 80–150 characters | Condiment marketing content is mostly short text. This range balances semantic completeness and chunking granularity, avoiding excessive splitting that breaks scene associations |
Incremental Index Switch | Enabled | Marketing content update frequency fluctuates widely. Incremental indexing avoids excessive time spent on full reconstruction, adapting to mixed scenarios of batch and single-item updates |
Embedding Model Selection | Open-source models adapted for short text (such as bge-small-zh) | Most condiment marketing content is short text. This type of model has higher short text encoding accuracy, improving recall matching performance |
Recall count | Top 3–5 entries | Single marketing content is clearly tied to specific scenarios. Too many recalls will introduce irrelevant information, while too few will fail to cover full requirements |
PARSE_FILE_TIMEOUT_SECONDS | 300 seconds | Long documents uploaded in batches (such as live streaming scripts) take longer to parse. This duration covers parsing needs for most conventional documents |
Similarity threshold | 0.72–0.85 | It is necessary to distinguish marketing content of the same category with different specifications. This range balances recall precision and coverage |
The parameter values provided on this page are common starting points for configuration. Actual values are affected by material form, data volume, and business rules. Specific issues require specific analysis. It is recommended to test on your own samples before finalizing settings.
Three common mistakes
- Issue: An error occurs when uploading a CSV file after upgrading the version, with the prompt "Text chunking failed". Cause: The default chunking parameters are updated in the new version, and the original configuration for
Chunk sizeis not synchronized. This leads to a mismatch in the chunking logic for short marketing copy. - Issue: The knowledge base indexing task gets stuck, with the interface showing incomplete status and no progress updates. Cause: Batch uploaded marketing documents include unstructured transcribed ingredient list content, and parsing times out without triggering an automatic retry mechanism.
- Issue: The indexing completion status cannot be confirmed after calling the file upload API. Cause: The returned
task_idis not used to poll the/v1/query/taskinterface for verification, and a status callback is not configured.
How to confirm the configuration is set correctly
- Upload a single short marketing copy (such as a social media promotion message), check if the chunking results align with the expected chunk length, and verify that the
Chunk sizeconfiguration takes effect. - Submit a batch marketing document upload task, check if the indexing progress updates normally, and confirm that the
Incremental Index Switchis enabled and has no timeout exceptions. - After modifying the vector model, initiate a knowledge base recall test, compare the differences in recall results across different models, and confirm that the model switching logic takes effect.
- After calling the file upload API, use the returned
task_idto poll the corresponding interface, and verify that index completion status feedback can be obtained.
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.