What the Data for This Category Looks Like
Refined oil sales service consultation data mainly comes from gas station operation manuals, fuel price lists, invoice handling guidelines, complaint handling specifications, and fuel card business process documents. The data includes two categories: structured and unstructured. Structured data includes fields such as fuel type, unit price, effective time, and processing nodes, with units of yuan/liter, times, and pieces. Unstructured data consists of process descriptions and frequently asked questions. Price lists are updated in real time alongside price adjustments. Operation manuals are updated quarterly or per compliance requirements. Service guideline documents are updated regularly over the long term.
What Constraints These Characteristics Impose on Vector Models and Indexing
Structured data has fixed fields and frequent updates, so vector indexes must support filtering by field metadata to avoid recalling expired price information. Unstructured service guideline text is lengthy, so reasonable chunking is required to retain semantic integrity and prevent cross-chunk truncation of process descriptions. Short-text service consultation questions account for a high share of queries, so vector recall matching accuracy must adapt to short-text semantics. Near-real-time index updates are also needed to sync the latest price list data. Additionally, individual gas stations have minor differences in business processes. The index must support data isolation by station to avoid mixing consultation results across sites.
How to Set Configurations
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to structured fragments and unstructured process text from refined oil sales documents, ensuring semantic integrity |
| Overlap Length | 100–200 characters | Prevents truncation of price adjustment notes and service processes across chunks, ensuring coherent context |
| Similarity Threshold | 0.75–0.85 | Adapts to short-text queries for service consultations, filtering low-match expired documents or irrelevant processes |
| Single Search Limit | Top 8 entries | The number of associated documents for gas station service consultations is concentrated; excess entries increase context redundancy |
defaultToken | 4096 | Matches the maximum input token count of Qwen3-Embedding-8B, ensuring complete embedding of long-text service guidelines |
maxToken | 8192 | Covers the full context length requirement for spliced service process documents after knowledge base retrieval |
Three Common Misconfigurations
- Phenomenon: Connection timeout or 500 status code is returned when connecting to Qwen3-Embedding-8B deployed via VLLM. Cause: The VLLM service port is not opened to the network environment where FastGPT is deployed, or the model interface address entered in the configuration is incorrect.
- Phenomenon: Knowledge base disk usage statistics only display the original file size, excluding split file chunks and embedding vectors. Cause: FastGPT’s vector storage path is not configured correctly, resulting in embedding vectors not being included in total usage statistics.
- Phenomenon: The number of retrieval results does not match the set single search limit, and expired price information is not filtered. Cause: Metadata filtering rules are not configured for structured price list data, resulting in expired price adjustment documents being recalled.
How to Confirm Correct Configuration
- Upload a local fuel price list document, run a unit price query test for the corresponding fuel type, and check if the retrieval results include the corresponding fragment of this document.
- Access the knowledge base management page of FastGPT v4.17.0, view disk usage statistics, and confirm that the total usage of original files, split chunks, and embedding vectors is displayed.
- View the vector model call logs, confirm that requests for Qwen3-Embedding-8B are normally received and embedding vector results are returned.
- Modify the chunk length configuration, re-trigger document splitting, and check if the updated number of chunks is displayed in the knowledge base document list.
The values provided here are common starting points and should be evaluated against the reader's own samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-28.