What Data for This Category Looks Like
Data primarily originates from public corporate financing announcements, securities exchange disclosure documents, and first-hand information submitted by partner institutions. Updates are released daily, covering public financing updates from the current day and the prior 7 business days. Each document is a structured record with fields including full company name, financing round, financing amount, list of joint investors, disclosure date, and affiliated industry sector. Amount units use RMB ten thousand yuan uniformly. Date fields follow the YYYY-MM-DD format.
Constraints for Vector Models and Indexing
The daily incremental update requirement means the index must support high-frequency, small-batch writes. This avoids resource consumption from full index reconstruction. Structured fields include numeric amounts, enumerated financing rounds, and multi-value investor lists. Adapted vectorization logic must be designed for different field types. For example, unify amount units before converting to text embeddings. This avoids format incompatibility between numeric types and text embeddings. Single record text length is relatively fixed, but multi-value fields can extend overall length. Configure a reasonable text truncation threshold to prevent ultra-long fields from disrupting vector consistency.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
embedding_model | qwen-text-embedding-v3 or text-embedding-3-large | Supports combined vectorization of structured short text; models accessible via OneAPI reduce integration costs |
chunk_size | 800-1200 characters | The combined text length of a single financing daily report record mostly falls between 500-1000 characters. This range fully preserves field information and avoids excessive truncation |
index_batch_size | 50-100 records/batch | Daily new record volume is moderate. Small-batch writes reduce index write pressure and align with incremental update rhythms |
rerank_top_n | Top 3-5 results | Relevant results for financing daily reports need to accurately match user queries for rounds and amount ranges. A small number of reranked results meet retrieval needs and avoid interference from redundant information |
similarity_threshold | 0.75-0.85 | Semantic similarity thresholds for structured data must be higher than general text scenarios to filter low-match irrelevant financing records |
embedding_api_timeout | 30 seconds | Embedding model interfaces accessed via OneAPI typically have response delays between 10-25 seconds. This timeout setting avoids unnecessary waits while covering normal response durations |
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. It is recommended to test on your own samples before finalizing settings.
Three Common Configuration Mistakes
- Phenomenon: When an
embedding_modelis configured as a Tongyi multimodal vector model, the interface returns a 400 error with the prompt "OpenAI format calls are not supported". Cause: Tongyi multimodal vector models only support native SDK calls and do not support OpenAI-compatible format API requests. - Phenomenon: After enabling
rerank_modeland enabling reranking in index configuration, online recall test results are not sorted by rerank scores. Cause: The similarity threshold for rerank trigger was not configured in the retrieval module, or the rerank model's API key was not correctly bound. - Phenomenon: A timeout error occurs during full index reconstruction, with the log showing the
ETIMEDOUTstatus code. Cause:index_batch_sizeis set too large, exceeding the single-write concurrency limit of the vector database.
How to Confirm Configuration Is Complete
- Run a vectorization test for a single financing record. Verify that the returned vector dimension matches the official published parameters of the selected embedding model.
- Submit a test query containing the target company name and financing round. Verify that the sorting logic of the recall results conforms to the configured reranking rules.
- Simulate a single-batch incremental write process. Check that there are no timeouts or connection errors during index writing.
- View the embedding interface's response logs. Confirm that there are no 400 or 500 series abnormal errors.
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.