What the data for this category looks like
Data sources for film and theater marketing content include official publicity and distribution systems, theater schedule management systems, and cooperative media contribution channels. Update rhythm fluctuates with project cycles. Update frequency is higher during new film scheduling and pre-release publicity periods. Daily schedule adjustments are synced weekly. Document structure primarily uses structured fields paired with rich text, including unique film identifier, film title, release window, single session duration, marketing copy, poster resource URLs, and other fields. Units include minutes, session numbers, and similar values.
What constraints these characteristics impose on the knowledge base retrieval and recall workflow
Mixed structured and rich text document structure requires retrieval and recall to match both keywords and structured attributes, to avoid confusing marketing content for multiple films in the same release window. Fluctuating update rhythm requires incremental update logic to adapt to both high-frequency and low-frequency update scenarios, avoiding excessive resource usage from full updates. Marketing copy has strong timeliness and short individual content, requiring recall weights to prioritize timeliness and precise matching. The presence of resource fields such as posters requires retrieval results to associate complete resource addresses, avoiding returning only text summaries that fail to meet requirements for accessing marketing materials.
How to set configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
recall_top_k | Top 15 results | Film and theater marketing materials are mostly short text paired with structured fields. Sufficient candidate results must be covered before reranking to filter precise content |
similarity_threshold | 0.72–0.85 | Marketing copy has consistent style but large detail differences. This range avoids missing precise matches or recalling irrelevant materials |
chunk_size | 800–1200 characters | Individual marketing copy contains film information and event details. This length avoids breaking critical context while controlling per-segment retrieval load |
enable_incremental_update | Triggered according to publicity nodes | Adapts to the high-frequency update during new film publicity periods and weekly daily schedule sync rhythm, reducing unnecessary full synchronization overhead |
parse_image_path | Automatically parse and upload to hosted storage | Adapts to the localization requirement of embedded poster links in MD documents, solving the issue of external link failures |
search_filter_collections | Filter by film tags | Supports excluding marketing material collections from non-current projects to accurately recall target content |
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. It is recommended to test on your own samples before finalizing settings.
Three common mistakes
- Phenomenon: After uploading MD-format marketing materials to the knowledge base, embedded poster image links fail to load normally, and a 404 error is displayed on the interface. Cause: The
parse_image_pathparameter is not configured, and local or external image paths are not automatically replaced with stable resource addresses hosted by the knowledge base. - Phenomenon: Retrieval results only include text summaries, and complete links or poster addresses for original materials cannot be obtained. Cause: The
return_source_urlconfiguration item is not enabled, or resource address fields are not bound in the recall configuration. - Phenomenon: Retrieval results include marketing content for multiple different films, making it impossible to accurately match the target project. Cause: The
search_filter_collectionsparameter is not configured, and target collections are not filtered by film tags or unique identifiers.
How to verify correct configuration
- Upload a single MD test file containing a poster link, check whether images display normally in the content parsed by the knowledge base.
- Initiate a retrieval request for a single film, verify whether returned results only include marketing materials related to that film.
- Trigger an incremental update task, check whether the knowledge base only synchronizes newly added publicity materials without overwriting full historical data.
- View the metadata fields of retrieval results, confirm whether complete access addresses for original resources are included.
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.