What Data Looks Like for This Category
TCM marketing content data primarily comes from official TCM standard documents, pharmaceutical company self-developed product manuals, health and wellness service script templates developed by financial, insurance, and wealth management institutions, and compliant materials published by industry associations. Update rhythms adjust alongside official standard revisions, new product launches, and marketing customer acquisition campaigns, with no fixed uniform cycle. A single document typically includes four field types: basic product attributes, function descriptions, compliance reminders, and marketing copy tailored for target customer groups. Some documents require noting herbal processing specifications and dosage units such as grams and milliliters. Marketing script documents also include health service expression content targeted at financial or insurance clients.
What Constraints These Characteristics Impose on Knowledge Base Retrieval and Recall
TCM marketing content has dense professional terminology, including exclusive expressions such as medicinal properties, meridian tropism, and processing specifications. Scripts targeted at financial, insurance, and wealth management clients must balance professional compliance and customer adaptation, which increases matching errors for general word segmentation models. Targeted adjustments to word segmentation rules are required. No fixed update rhythm requires the knowledge base to support incremental content updates, avoiding resource consumption from full re-imports and adapting to the rapid adjustment needs of marketing campaigns. Single documents vary widely in length: short scripts are only a few hundred characters, while long compliance documents can reach thousands of characters. Flexible segmentation strategies must be set to avoid splitting professional term combinations. Some documents include dosage unit fields, which must be associated with unit information during retrieval matching to ensure the accuracy of recall results and avoid misleading clients.
How to Configure Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
chunk_size | 800–1200 characters | TCM content has dense professional terminology. This segmentation range preserves the integrity of most professional expressions and avoids splitting term combinations |
similarity_threshold | 0.72–0.8 | Professional terminology carries higher weight in TCM marketing content. This threshold filters irrelevant recall results while retaining relevant compliance documents |
recall_top_k | Top 6–8 results | Single TCM documents have high professional information density. Too many recall results will exceed the context window limit, while too few will fail to cover complete marketing scenarios |
UPLOAD_INCREMENTAL_SUPPORT | Enabled | TCM content updates have no fixed cycle. Incremental uploads only update modified documents, reducing resource usage |
PARSE_SPECIAL_TERM | Enabled | TCM contains a large number of exclusive professional terms. Enabling this setting preserves complete word segmentation of terms and improves matching accuracy |
MAX_DOCUMENT_SIZE | 20000 characters | Covers the length limit of most long compliance documents, avoiding parsing timeouts for individual documents |
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 individual analysis, and it is recommended to test on your own samples before finalizing settings.
Three Common Configuration Mistakes
- Incremental update task completes successfully but old documents are not overwritten. Phenomenon: Old versions of marketing scripts remain in the knowledge base, and update logs show the incremental upload task status is successful. Cause: The
UPLOAD_INCREMENTAL_SUPPORTconfiguration item is not enabled. The system performs full overwrite by default and does not execute incremental updates. - Extended AI response time following knowledge base retrieval. Phenomenon: System monitoring shows normal retrieval phase duration, but the response generation phase takes more than 30 seconds. Cause: The total character count of recalled documents exceeds the context window limit set by the
maxContextconfiguration, requiring additional processing of redundant content. - Custom embedding model call failure. Phenomenon: Embedding status displays failure after knowledge base parsing, with error code
400 Bad Requestreturned. Cause: TheEMBEDDING_API_ENDPOINTparameter is not correctly configured to point to the locally deployed model service address, or required request authentication information is not added.
How to Verify Proper Configuration
- Upload a test TCM marketing document, review parsed segmentation results to confirm professional terms are not split and segmentation length aligns with configured requirements.
- Initiate a retrieval targeting TCM professional terminology, verify the number of returned recall results matches the configured recall top k value, and that similarity scores meet preset standards.
- Submit a modified document for incremental upload, wait for task completion, compare the update time of the document in the knowledge base with its modification time to confirm content has been synchronized.
- Use the system-provided embedding model testing tool to confirm the request and response formats of the custom embedding model meet system requirements.
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.