Skip to main content
XpertAI Knowledge Base turns enterprise documents, web pages, and business material into trusted knowledge that digital experts and workflows can use. It handles content extraction, chunking, embedding, structured filtering, retrieval, and citation so agent answers remain grounded in traceable source material. A knowledge base can be exposed as a tool that an agent calls autonomously or as an explicit retrieval step in a workflow. One knowledge base can serve several business domains, while fixed and agent-generated filters on each agent-to-knowledge-base binding control the scope used for a particular application.

​
Core capabilities

​
From document to grounded answer

  1. Ingest content: upload files, scrape web pages, or connect business document sources through a knowledge pipeline.
  2. Parse and chunk: extract and clean text, then create chunks suitable for retrieval.
  3. Embed and index: use the configured embedding model and write vectors to the vector database.
  4. Constrain candidates: apply access boundaries, enabled-content checks, fixed filters, and agent-generated filters.
  5. Explore relationships when useful: for GraphRAG-enabled knowledge bases, let the agent discover entities and relations before forming the final retrieval query.
  6. Rank semantically: rank the remaining chunks by vector similarity and apply TopK and score thresholds.
  7. Generate with citations: pass the results to downstream models while preserving traceable document sources.
RAG indexing

​
PDF processing and source linking

For scans, engineering manuals, contracts, and complex table PDFs, the Baidu PaddleOCR‑VL document converter can preserve paged layouts, tables, images, coordinates, and source maps. After processing, document details provide a side-by-side analysis preview: the original PDF and layout bounds appear on the left, while Markdown, structure, tables, images, and JSON appear on the right. Scrolling, active pages, and selected analysis blocks synchronize in both directions. Baidu Unlimited‑OCR focuses on long-document Markdown extraction. MinerU supports both its official Precise Parsing API and self-hosted services, with OCR, formula, table, and VLM/pipeline options. The Pdfium PDF to Markdown tool lets an agent process a task attachment temporarily. See PDF Processing and Linked Analysis Preview to choose and configure the appropriate capability.

​
Intelligent filtering for multi-domain knowledge

When engineering, water-resources, logistics, and other material share one knowledge base, unrestricted semantic matching may retrieve similar but inapplicable content. XpertAI can apply structured conditions before vector search:
  • administrators define fixed filters for a domain, project, folder, or document-status boundary that the agent cannot change;
  • when Allow agent automatic filtering is enabled, the agent can infer a year, file format, region, or keyword from the question and add request-specific conditions;
  • both layers are always combined with AND, so the agent can narrow but never bypass the fixed scope;
  • if no reliable dynamic condition can be inferred, the agent omits it and still searches within the fixed scope.
For example, an administrator can constrain a quoting agent to “water resources + effective documents.” For “2025 PDF rate schedules,” the agent can additionally filter for file extension = pdf, effective year = 2025, and file name contains rate schedule. See Intelligent Retrieval Filtering for fields, operators, configuration steps, runtime behavior, and the support matrix.

​
Document and metadata management

XpertAI automatically maintains these system attributes:
  • file name and logical knowledge base folder path;
  • file extension and MIME type;
  • document category and source type;
  • creation and update timestamps.
Teams can define business fields in the metadata schema and supply typed values during upload, document editing, chunk editing, or bulk maintenance. Document-scoped fields apply to the whole document; chunk-scoped fields distinguish sections or content categories inside a document. Renaming or moving a file and updating metadata refresh retrieval attributes without regenerating vectors, so day-to-day classification changes take effect quickly. See Maintaining Documents.

​
Ways to retrieve knowledge

See Ways to Use the Knowledge Base and Knowledgebase Workbench.

​
Testing and observability

Recall testing validates queries, filters, TopK, and similarity thresholds before release. Debug views distinguish fixed, test, and agent-generated filters and show the effective filter, candidate documents, candidate chunks, final hits, backend, and timing. Retrieval log access follows knowledge base permissions, and logs do not contain document body text. Normal agent answers show only content and citations by default, not internal filtering decisions. See Recall Test.

​
Current support

Intelligent filtering supports Vector, Graph, and Hybrid retrieval. PGVector and Milvus are the supported vector branches; Milvus Server 2.6.2 or later is required. Graph applies the filter to seed eligibility, relationship expansion, and evidence chunks. Hybrid filters both branches before merge and can fall back only to its already-filtered Vector branch. Chroma and Weaviate still return an explicit unsupported error for new vector filter conditions; XpertAI never silently degrades to an unrestricted search.
  1. Maintain documents and metadata
  2. Process and preview PDFs
  3. Build and use Knowledge Graph and GraphRAG
  4. Configure intelligent retrieval filtering
  5. Test recall
  6. Choose a knowledge base usage pattern
  7. Create a knowledge base through a pipeline