What Data for This Category Looks Like
Refinery and chemical production policy and regulatory basis data comes from oil and petrochemical industry safety specifications and production license requirements published by the National Energy Administration and the Ministry of Emergency Management, as well as internal production management rules and compliance manuals from affiliated state-owned enterprises. Data receives at least one annual update alongside national policy adjustments. Internal rules receive quarterly updates alongside production process optimizations. Most documents are PDF or Word files with document numbers and effective dates. These files include fields such as policy number, applicable production stage, specific clause content, and responsible executive department. Clauses are hierarchically marked by article, paragraph, and sub-paragraph. This page was created for FastGPT v4.17.0.
How These Characteristics Impact Knowledge Base Retrieval and Recall
Policy document effective dates and document numbers require precise matching of valid versions during retrieval. This prevents recalling expired clauses. Some compliance documents are scanned files. Plain text extraction from these files may produce garbled text or missing content. An additional OCR preprocessing step is required for these files. The hierarchical clause structure (by article, paragraph, sub-paragraph) requires retrieval and recall to retain contextual connections. This avoids losing clause hierarchy logic after splitting. Differences between multi-source data require distinguishing applicable scenarios between official policies and internal rules during recall. This prevents confusing execution standards.
Configuration Settings
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Refinery policy text uses formal, precise regulatory language. This range balances retrieval precision and coverage, avoiding irrelevant clauses or missed valid compliance references |
limit | Top 8–12 results | Single-round policy retrieval must cover enough clauses to match complex production scenarios, while avoiding excessive redundant information that disrupts decision-making |
usingReRank | Enabled, retain top 5 reranked results | Policy text has high semantic similarity. Reranking filters clauses that match literally but do not fit actual use cases, improving retrieval precision |
datasetSearchUsingExtensionQuery | Enabled | Policy questions often include specialized terminology. Query expansion supplements synonyms and industry terms to cover more relevant clauses |
searchMode | vector | Refinery policies have strong semantic connections. Vector retrieval accurately matches specialized terminology, and is better suited for semantic queries on regulatory text than keyword retrieval |
collectionIds | Bind exclusive policy knowledge base collection | Refinery production policy data is fully isolated from other categories. Specifying an exclusive collection ID prevents recalling unrelated documents |
Three Common Mistakes
- The interface returns the error
invalid configuration parameter name "hnsw.max_scan_tuple. This occurs when using parameters outside the FastGPT official supported list, instead of using specified configuration items for setup. - Excessive knowledge base search latency occurs. This happens when no reasonable
limitparameter is set, recalling too many redundant clauses that increase retrieval computation load, or when reranking functionality is not enabled, which amplifies processing overhead from invalid results. - Garbled text or no matching content appears in search results for scanned PDF policy documents. This occurs when OCR preprocessing is not performed on scanned files. Directly extracted image text cannot be properly indexed, leading to matching failures.
How to Verify Proper Configuration
- Review the parameter list on the configuration page, confirm only FastGPT officially supported configuration items are used, with no custom non-supported parameters.
- Run a single-round policy retrieval test, verify the number of returned results matches the
limitparameter setting, and that results include valid policy clauses. - Upload a scanned policy document, perform OCR preprocessing, then confirm the indexed result has no garbled text and can be properly matched during retrieval.
- Run a batch multi-question retrieval test, confirm each question’s retrieval results are isolated independently, with no result interference across questions.
The values provided here are common starting points, and should be evaluated against your own specific 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.