What the data for this category looks like
Oil and gas exploration data primarily comes from exploration geological reports, real-time drilling logs, reservoir dynamic monitoring reports, production progress ledgers, and industry technical standard documents. Structured data includes fields such as well ID, daily oil production, formation pressure, with units mostly being barrels, cubic meters, and megapascals. Unstructured data mostly consists of oil test analysis and geological modeling reports ranging from thousands to tens of thousands of characters per document. Data update frequencies vary widely: real-time drilling logs are updated minute-by-minute, monthly exploration reports are updated per calendar month, and industry standard documents are revised every 1 to 2 years.
What constraints do these characteristics impose on vector models and indexing?
Standardized units and field names for structured data require vector models to accurately recognize professional terminology and numerical associations, to avoid matching failures caused by unit conversion errors. The high proportion of long documents requires indexing to support retaining contextual associations after segmentation, to avoid breaking geological analysis logic. Real-time data updated minute-by-minute requires indexing to support incremental refreshes, to avoid excessive time spent on full reconstruction. Mixed storage of multi-source heterogeneous data requires index structures to be compatible with both structured metadata and unstructured text vectors, to enable precise associated retrieval.
How to set the configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
embedding_model | Prioritize bge-m3 or text-embedding-v3 | Adapt to the vector encoding accuracy of professional terminology in the oil and gas exploration field, and support long text segment encoding |
chunk_size | 800–1200 characters | Balance the semantic integrity of a single text segment and retrieval recall density, avoid excessive length leading to semantic dilution or excessive short length leading to loss of professional associations |
chunk_overlap | 100–150 characters | Retain the context of geological terminology in adjacent segments, avoid professional logic in drilling logs and oil test reports being broken by segmentation |
retrieval_top_k | Top 10–15 results | Cover the retrieval scope of multi-source heterogeneous data, adapt to the scattered characteristics of oil and gas exploration data |
rerank_top_k | Top 5–8 results | Filter redundant recall results, improve retrieval efficiency while retaining core professional information |
index_refresh_interval | 60 seconds | Adapt to minute-by-minute updated real-time drilling data, achieve a balance of incremental index refresh |
The parameter values provided on this page are conventional recommendations used as a starting point for configuration. Actual values are affected by material form, data volume, and business rules. Specific issues require case-by-case analysis, and it is recommended to test on your own samples before finalizing the settings.
Three common configuration mistakes
- Symptom: When configuring
bge-m3deployed viaollamaas the vector model, the interface prompts a vector service connection failure, or returns empty vector results. Cause: The local port mapping of ollama was not correctly configured in FastGPT, or an attempt was made to connect before the vector model finished starting. - Symptom: When configuring
text-embedding-v3, the interface prompts "No available channels under the current default group". Cause: The corresponding vector retrieval service permission was not activated for the current group, or the API key was not bound to the corresponding resource group. - Symptom: Knowledge base question and answer response times are too long, with significant delays in the post-retrieval reranking stage. Cause: The
chunk_sizeis set too large, leading to excessively high vector dimensions for a single text segment, or thererank_top_kvalue exceeds a reasonable range, increasing reranking calculation load.
How to confirm the configuration is successful
- Upload a single drilling log document, confirm that there are no errors in the vector generation process, and the interface shows that vector generation is completed.
- Enter professional oil and gas exploration retrieval terms, verify that recalled results include matching fields, units, and source document details.
- Batch upload multiple documents of different types, confirm that there is no persistent backlog in index refresh status.
- After adjusting the vector model configuration, verify that the connection status shows "Connected", with no errors related to channel permissions.
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.