What the data for this category looks like
Data sources for film and theater include cinema monthly operation reports, film distribution agreements, scheduling plans, box office settlement documents, passenger flow statistics ledgers, and similar materials. Update rhythms vary: scheduling plans update 1 to 7 days in advance, box office data updates after daily settlement, and copyright contracts and operation reports update after signing or the end of a monthly cycle. Single documents are mostly multi-page structured text, with fields including cinema ID, screening sessions, ticket price per ticket, box office revenue, passenger flow count, copyright holder name, and more. Units include sessions, yuan per ticket, ten thousand yuan, passenger trips, and others.
What constraints these characteristics impose on the knowledge base retrieval and recall link
Multi-source data with varying update rhythms requires the retrieval link to distinguish between static and dynamic data, to avoid recalling expired box office or scheduling information. The large number of structured fields and inconsistent units requires field normalization before retrieval, to ensure fields with the same business meaning are matched uniformly. The long length of single documents requires retaining field associations during segmented retrieval, to avoid breaking the complete business context of a single cinema or film, while reasonably controlling segment length to balance retrieval accuracy and model processing efficiency.
How to set configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
chunk_size | 800–1200 characters | Film and theater documents are mostly long-form operation reports and contracts. This segment range retains the complete business context of a single cinema or film, avoiding broken field associations |
recall_top_k | Top 8–12 results | Film and theater due diligence requires covering multi-dimensional data including scheduling, box office, and copyright. Too few recalls will miss key information, while too many will increase model processing load |
similarity_threshold | 0.72–0.85 | Most fields in film and theater scenarios are precise business terms. A threshold that is too low will introduce irrelevant documents, while a threshold that is too high will filter out some relevant marginal business data |
rerank_top_n | Top 3–5 results | Due diligence reports require prioritizing core information. Re-ranking retains highly relevant results, preventing models from being distracted by low-value content |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Monthly cinema operation reports are mostly multi-page PDF documents. This duration covers the complete parsing and field extraction process |
knowledgeSearch | Dynamically pass cinema ID and film name | Due diligence reports require retrieval targeting specific cinemas or films. Dynamic parameter passing can accurately locate target data sources, improving retrieval targeting |
The parameter values provided on this page are common recommended starting points for configuration. Actual values are affected by material format, data volume and business rules. Specific issues require specific analysis, and it is recommended to test on your own samples before finalizing.
Three common errors
- Phenomenon: Retrieval results only include text datasets, and documents in other formats are not recalled. Cause: The parsing switch for the corresponding format files is not enabled, or the parsing configuration does not cover field extraction rules for non-text formats.
- Phenomenon: The AI response does not reference knowledge base content associated with the dynamically passed
knowledgeSearchvariable. Cause: TheknowledgeSearchvariable is not bound to the query parameters of the knowledge base retrieval node in the workflow, or the variable transfer format does not meet interface requirements. - Phenomenon: Conflicting content appears in recall results, and no automatic analysis prompt is triggered. Cause: The knowledge base content conflict detection configuration is not enabled, or the conflict detection threshold is set too high to recognize field value differences.
How to confirm the configuration is correctly set
- Upload a film and theater monthly operation report document, trigger knowledge base parsing, and check if the parsed data fields include preset business fields.
- Manually enter a query statement targeting a specific cinema, and check if the number of retrieved results matches the configured value range.
- Pass the
knowledgeSearchvariable for dynamic retrieval, and check if the reference sources of the returned results include the metadata of the corresponding document. - Upload two documents with field conflicts, and check if the system triggers a content conflict prompt, or if the retrieval result automatically filters conflicting content.
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.