Vector Models and Indexing for Traditional Chinese Medicine

What data for this category looks like

Data sources for Traditional Chinese Medicine (TCM) marketing content cover four categories: official pharmaceutical company product materials, industry processing specifications, academic clinical documents, and marketing promotional materials. Update cycles are split into bulk updates and daily iterations. Bulk updates trigger when new products launch or industry standards are revised. Daily iterations occur alongside marketing materials, live stream scripts, official account posts, and similar content. Document structures fall into two categories: structured parameter tables and non-long text. Structured fields include generic drug names, brand names, nature, flavor and meridian tropism, functions and indications, usage and dosage, contraindications, and others. Units include grams, milliliters, daily dosage, and similar units. Unstructured content includes product popular science text, marketing scripts, and similar materials.

What constraints these characteristics impose on vector models and indexing

TCM marketing content has both structured parameters and unstructured text. This requires vector indexes to support both field-level semantic encoding and full-text semantic retrieval. Dense volumes of TCM specialized terminology in the content require vector models adapted for domain semantic understanding. This avoids semantic deviations from generic models when processing professional expressions. The coexistence of bulk updates and daily iterations requires indexes to support both incremental refresh and full reconstruction modes. This balances real-time performance and index system load. Multi-dimensional business classifications such as category, dosage form, and target population require index metadata to support multi-label classification. This facilitates subsequent retrieval and export by business dimensions.

Configuration Settings

Config Item	Suggested Value	Rationale
`chunk_size`	800–1200 characters	Balances semantic integrity of TCM specialized terminology and complete extraction of short marketing copy, avoiding segmentation that cuts off continuous professional expressions
`chunk_overlap`	100–150 characters	Preserves contextual continuity between segments, preventing truncation of continuous professional phrases such as "nature, flavor and meridian tropism" and "functions and indications"
`retrieval_top_k`	Top 6–10 results	TCM marketing content has strong professional relevance. Too many retrieved results introduce irrelevant information, while too few fail to cover complete requirement scenarios
`similarity_threshold`	0.72–0.80	Distinguishes semantic similarity between similar TCM products, avoiding confusion between drugs with similar efficacy
`index_incremental_refresh_interval`	Once per hour	Adapts to the daily iteration cycle of marketing content, balancing real-time performance and index system load
`parse_file_max_size`	200 MB	Adapts to bulk import requirements for large-volume documents such as TCM industry processing specifications and academic papers

The parameter values provided on this page are common starting points for configuration. Actual values are affected by material format, data volume, and business rules. Specific issues require case-by-case analysis, and it is recommended to test on your own samples before finalizing settings.

Three common mistakes

Issue: After importing Excel files from a folder, the functions and indications field is empty in the parsed results. Cause: Structured table field extraction configuration is not enabled, and mapping rules are not set for TCM-specific business fields.
Issue: When deploying an external vector model on an ARM soft router, frequent 504 Gateway Timeout errors occur. Cause: The soft router's single-core computing power is insufficient to support batch inference of the vector model, and the interface timeout threshold is set too low.
Issue: After bulk updating TCM marketing content, newly imported documents are not included in the vector index. Cause: The index_incremental_refresh_interval parameter is not configured, and only full index refresh is relied on, which cannot synchronize incremental data in a timely manner.

How to confirm proper configuration

Upload a single TCM product instruction manual. Check that the parsed segments include complete nature, flavor and meridian tropism and functions and indications fields, and confirm that segment lengths match the preset rules.
Initiate a test query for TCM efficacy. Verify that the number of retrieved results matches the retrieval_top_k setting, and confirm that index refresh has taken effect.
View the vector model's inference logs. Confirm that the response delay of the external interface meets business requirements, and no timeout errors occur.
Export a portion of index data. Check that classification filtering is available by TCM category and document type, and confirm that metadata configuration is correct.

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.