Vector Models and Indexing for State-owned Large Bank

What the Data for This Category Looks Like

Data for state-owned large bank financing daily reports comes from internal credit ledgers, publicly disclosed data from the National Interbank Funding Center, and selected content from the central bank’s financial market operations. Data collection and generation for the current day finishes at 17:00 daily. Retroactive correction of same-day business data is supported. Documents use structured table format, with fields including full financing entity name, financing amount, financing term, weighted average interest rate, approval completion time, and peer benchmark data. Financing amount uses the unit of RMB 100 million yuan. Financing term uses the unit of calendar days. Weighted average interest rate uses the unit of percentage. The number of business entries per document fluctuates based on daily credit transaction volume.

Constraints Imposed on Vector Models and Indexing

Multiple structured fields exist, including both numeric and text fields. Vector models need to support mixed-type feature extraction, to avoid losing numeric business information when using a single text vector model. Updates run at a fixed daily time, and same-day data correction is supported. Indexes need to support low-latency incremental writes and rollback mechanisms, to ensure consistency of data updates. Each daily report contains tens to hundreds of business records. Precise chunking by individual business entries must be supported, to avoid splitting entire documents and prevent cross-business context confusion. Data involves financial sensitive information. Vector indexes need to support end-to-end encrypted storage, to comply with financial data regulatory requirements.

Configuration Settings

Configuration Item	Recommended Value	Rationale
`embedding_model_endpoint`	`http://localhost:11434/v1`	Compatible with OpenAI interface format for Qwen3-Embedding-8B:F16 deployment
`chunk_size`	`800–1200 characters`	Single business description in state-owned large bank financing daily reports is approximately 500–800 characters long. This range fully covers core information and avoids redundancy
`embedding_batch_size`	`32–64 items/batch`	Matches GPU memory usage for Qwen3-Embedding-8B:F16, balances inference performance and hardware resource consumption
`milvus_collection_shards`	`2–4 shards`	Matches daily vector data write volume, balances write performance and cluster resource overhead
`rerank_model_endpoint`	`http://localhost:11435/v1`	Compatible with deployment interface format for Qwen3-Rerank
`vector_search_topk`	`Top 15–20 results`	Covers potential related businesses per daily report, avoids computational overload

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. Testing should be completed on relevant samples prior to final configuration.

Three Common Configuration Mistakes

Issue: 500 Internal Server Error or connection refused errors occur when connecting a custom embedding or reranking model. Cause: Model endpoint IP, port, or interface path is configured incorrectly, or the locally deployed model service has not enabled access permissions for the corresponding port.
Issue: Milvus container fails to start, with logs indicating PostgreSQL service startup timeout. Cause: A full Milvus Compose file with additional database dependencies is used directly. Standalone vector indexes do not require mounting PostgreSQL storage.
Issue: Index recall results are abnormal, or out of memory errors occur. Cause: The configured embedding_batch_size or chunk_size exceeds the current hardware's carrying capacity, leading to model inference or index write failure.

How to Verify Proper Configuration

A single financing daily report business description is entered into the FastGPT model test page, and the vector dimension returned by the embedding model is verified to match the configured vector_dimension parameter.
A test state-owned large bank financing daily report document is uploaded, parsed chunked content is reviewed, and the chunk length is confirmed to fall within the preset chunk_size range.
The Milvus container is started, running logs are checked, no dependent service errors are confirmed, and the target index collection is verified to be created successfully.
A simulated recall request is sent, the number of returned results is verified to match the configured vector_search_topk parameter, and no dimension mismatch related errors are present.

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.