• Knowledge Ingestion — Feed in PDFs, documents, or data from external APIs. Content is automatically processed and stored in a searchable knowledge base — upload via API or directly through the admin chat.
• Batch Processing — Ingest multiple sources at once. The system processes each document individually and returns per-item status reports.
• Semantic Search — Ask a question, get the most relevant passages from your knowledge base — ranked by relevance and filterable by metadata.
• Knowledge Chat — An AI assistant that searches your knowledge base autonomously. It decides when to search, what to look for, and can refine its queries multiple times before delivering an answer. Available through a built-in UI or via API for integration into existing systems.
• Knowledge Base Administration — A separate admin chat lets you manage everything through conversation: add content, search, list entries, inspect details, and delete — all in natural language.

• Vector Database with @operaide/vector — aktorVektorDatabase initializes a AgentDB-backed vector store. The same database client is shared across all five Reaktors — embedding, retrieval, chat, admin, and batch processing all read from and write to the same store.
• Embedding Pipeline — aktorVektorEmbeddingPipeline handles the full ingestion flow: content parsing (e.g. PDF via @operaide/document), text chunking with configurable overlap, embedding generation (OpenAI text-embedding-3-small by default), and batch storage. A custom aktorBase64ToFile converts uploaded base64 data URLs into proper File objects before they enter the pipeline.
• Retrieval as a Tool — aktorToTool wraps aktorVektorRetrievalPipeline into an AI-callable tool with a Zod-validated parameter schema. The database client is injected as a dependency — invisible to the AI, which only sees the query parameter. This pattern cleanly separates what the AI controls from what the system provides.
• Multi-Step Tool Orchestration — aktorAICall is configured with max_steps: 5, allowing the AI to call tools multiple times per turn. It can search, review results, refine its query, and search again before generating a final response.
• Stream Processing for Batch Uploads — aktorForEachStream creates a stream from an array of files, aktorVar holds the current file, and aktorStreamProcessor runs the embedding pipeline for each item. This pattern enables efficient multi-file processing without loading all files into memory simultaneously.
• Composable Toolsets — Individual tools are created with aktorToTool, grouped with aktorToolSet, and merged with aktorCombineToolSets. The admin chat combines a file management toolset with a search/upload toolset into a single unified set the AI can use.

The Knowledge Chat system is composed of five Reaktors sharing a single vector database. The embedding Reaktor and its batch variant handle content ingestion — input enters aktorBase64ToFile, flows through aktorVektorEmbeddingPipeline (parsing, chunking, embedding, storage), and returns processing stats. The retrieval Reaktor exposes aktorVektorRetrievalPipeline directly for programmatic semantic search. The knowledge-chat Reaktor wraps the retrieval pipeline as a tool via aktorToTool and hands it to aktorAICall with max_steps: 5 — the AI decides autonomously when and how to search. The admin Reaktor is the most complex: it extracts files from chat messages with aktorFilesFromLastMessageAsFiles, sets up a stream processor for batch embedding, creates five tools (process, search, list, inspect, delete), merges them with aktorCombineToolSets, and gives the AI full knowledge base management capabilities through conversation.