What the data for this category looks like
This category’s marketing content data sources include structured store menus, member consumption tags, in-store time slot statistics, holiday event materials, user review snippets, and more. Update rhythms vary significantly: menus and fixed packages are updated quarterly or per event nodes, member tags and real-time consumption data are synced daily, and temporary promotion copy can be added at any time. Document structures cover structured fields such as sku_id, store id, per capita consumption, semi-structured event copy, unstructured user feedback snippets. Field units include yuan, number of visits, time slots, and more.
What constraints these characteristics impose on deployment and upgrade
A large number of structured fields with varying update rhythms require support for bulk import and incremental sync configuration of structured data during deployment. This adapts to different sync needs such as menu updates and member tag refreshes. A high proportion of long-text event copy and multi-dish combination introductions require adjusting context segmentation and recall parameters during upgrades. This adapts to parsing and recall of long content. Clear multi-store data isolation requirements require configuring multi-tenant permission parameters during deployment. This prevents mixing of marketing content from different stores. A large number of proper nouns such as dish names and store addresses require supporting custom TTS lexicon configuration during upgrades. This ensures accuracy of voice broadcasts.
How to set the configurations
These configurations apply to FastGPT V4.8.17 and later versions.
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
maxContext | 8000–12000 characters | Hotel and catering marketing content includes long event copy and multi-dish combination introductions, adapting to long-context recall needs |
PARSE_FILE_TIMEOUT_SECONDS | 300 seconds | Menu PDFs and event poster long images take longer to parse, avoiding timeout interruptions |
RECALL_COUNT | Top 6–8 entries | Associated knowledge base entries for a single store’s marketing content are limited, reasonably controlling recall volume to avoid redundancy |
SIMILARITY_THRESHOLD | 0.65–0.75 | Filters low-relevance dish or event content, ensuring accuracy of marketing responses |
UPLOAD_FILE_MAX_SIZE | 500 MB | Supports uploading high-definition material files for bulk menus and store events |
TTS_CUSTOM_ENDPOINT | Interface address of locally docker-deployed chatTTS | Adapts to pronunciation needs of proper nouns such as hotel catering dish names and store addresses |
The parameter values provided on this page are common recommended starting points for configuration. Actual values are affected by material form, data volume and business rules. Specific issues require case-by-case analysis, and it is recommended to test on your own samples before finalizing.
Three common errors
- Phenomenon: Field empty errors occur when inserting structured menu data after deployment. Cause: No structured data field mapping rules are configured, so the system cannot recognize imported fields such as store ID and dish price.
- Phenomenon: Off-topic outputs occur when processing long event planning copy. Cause: The
maxContextparameter value is too small to fully load long-text context, or long-text segment recall configuration is not enabled. - Phenomenon: The docker-deployed service cannot trigger custom TTS playback. Cause:
TTS_CUSTOM_ENDPOINTis not correctly configured to point to the interface address of the local chatTTS container, or container port mapping is not opened.
How to confirm the configuration is complete
- Upload a store menu PDF, check that the parsed fields include expected content such as dish name, price, store ID, and confirm that the parsing timeout and file size configurations take effect.
- Initiate a query containing long event copy, check that the number and similarity of retrieved context entries meet the preset range.
- Call the TTS interface with proper nouns such as "Buddha Jumps Over the Wall" and "Hubin Road Store", check that the pronunciation meets expectations, and confirm that the custom endpoint configuration is correct.
- Insert a structured member consumption data entry, check that the data is successfully stored in the database, and confirm that the field mapping configuration takes effect.
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.