What this category of data looks like
Data sources for TCM research reports include authoritative TCM industry journals, public documents from the National Medical Products Administration, annual R&D reports of TCM enterprises, and special survey materials from TCM research institutes. Three update cadences apply:
- Industry general research reports are updated quarterly and semi-annually.
- Policy-related research reports update synchronously with regulatory document releases.
- Enterprise in-house R&D data updates in real time alongside R&D progress.
Single documents include these modules:
- Variety origin
- Nature, flavor and meridian tropism
- Functions and indications
- Dosage and administration
- Quality control parameters
- Clinical research data
- Analysis of pharmacologically active components
Document fields include common name, Latin name, specification, content unit, production batch number, and other related items.
Constraints on Vector Models and Indexing
TCM research reports mix structured professional parameters and unstructured text. This requires vector models to support multi-field associated vectorization, to avoid semantic loss of structured data. Coexisting multiple update cadences requires index architectures to support flexible switching between incremental updates and full reconstruction. This adapts to different update needs for regular industry reports and real-time policy documents. Dense professional terminology with fixed expression patterns requires vector models to adapt to TCM domain terminology. This ensures consistent vector spaces for equivalent terms. Wide variation in document length, from hundreds of characters of variety descriptions to ten-thousand-character pharmacological studies, requires index segmentation strategies to support adaptive adjustment. This avoids excessive truncation or redundant segmentation.
How to configure the settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
embedding_model | Zhipu Embedding-3, or a general vector model fine-tuned for the TCM domain | Adapts to semantic representation of large volumes of professional terminology in TCM research reports, reduces vector space deviation for equivalent terms |
segment_length | 800–1200 characters | Adapts to the length distribution of TCM research reports, which include both short professional fields and long research texts, and avoids truncation of core information |
segment_overlap_rate | 100–150 characters | Retains contextual association between segments, avoids semantic breakage caused by cutting of professional terminology |
index_type | HNSW index | Balances retrieval speed and recall accuracy, adapts to the demand for batch retrieval of multiple documents in TCM research report knowledge bases |
recall_count | Top 8–12 results | Covers multi-dimensional professional information in TCM research reports, avoids omission of key parameters or clinical data |
similarity_threshold | 0.75–0.85 | Filters low-correlation non-professional matching results, ensures professional matching degree between retrieval results and TCM research reports |
The parameter values provided on this page are general recommendations for establishing configuration baselines. Actual values are affected by material form, data volume, and business rules. Specific issues require individual analysis. It is recommended to test on your own samples before finalizing settings.
Three common configuration errors
- Symptom: After calling the
create_train_orderAPI, the knowledge base retrieval results do not show newly added TCM research report data. Cause: The association between the data collection and training order is not bound, or the automatic vectorization switch is not enabled, resulting in failure to complete index construction for newly added data. - Symptom: Specified models such as Zhipu Embedding-3 cannot be found in the vector model configuration list. Cause: The API whitelist for the corresponding model is not added in the platform configuration file, or the model version is not compatible with the platform's vector model loading logic.
- Symptom: In a local deployment environment, the vector model can be pinged, but a
500 Internal Server Erroroccurs during retrieval. Cause: The access token permissions for the vector model are not configured correctly, or the communication port between the platform and the vector model is not open, resulting in failure of normal request routing.
How to verify successful configuration
- Call the vector model test API, input a piece of TCM professional terminology, and check that the returned vector results and the vector similarity of equivalent terms fall within the expected configuration range.
- Upload a test TCM research report document, check the index construction log, and confirm that parameters such as segment length and overlap rate match the configured items.
- Initiate a retrieval request, check that the number of recalled results matches the configured
recall_countparameter, and that similarity scores fall within the set threshold range. - Upload a TCM ingredient atlas image, initiate cross-modal retrieval, and confirm that the returned text research report content matches the image to the expected degree.
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.