Agentset

Executes vector-based semantic search across ingested documents combined with BM25 keyword matching, then applies a reranking algorithm to surface most relevant results. The system converts user queries to embeddings, searches a vector database (Pinecone or Qdrant), retrieves candidate documents, and reranks them using a learned-to-rank model before returning cited sources. This hybrid approach balances semantic understanding with keyword precision.

Enables answering questions that require retrieving and reasoning across multiple documents sequentially. The system performs iterative retrieval: initial query retrieves relevant documents, LLM generates follow-up queries based on retrieved context, system retrieves additional documents, and final answer synthesizes information across all retrieved sources. This is benchmarked on MultiHopQA, indicating support for 2-3 hop reasoning chains.

Provides webhook callbacks for document ingestion lifecycle events (started, completed, failed), enabling external systems to track ingestion status and trigger downstream workflows. The system sends HTTP POST requests to configured webhook URLs with event metadata (document ID, status, error details), allowing asynchronous monitoring without polling the API.

Enables enterprise customers to deploy Agentset in their own cloud infrastructure (AWS, Azure, GCP) or on-premise data centers, maintaining full data sovereignty and control. The deployment includes all components (API, vector database, LLM integration) and can be configured for high availability and disaster recovery. Data never leaves the customer's infrastructure.

Uses a consumption-based pricing model where customers pay per document page ingested ($0.01/page on Pro tier after 10,000 included pages) but have unlimited retrieval queries. This decouples ingestion costs from query volume, making the service cost-predictable for high-query-volume use cases. Free tier includes 1,000 pages and 10,000 retrievals/month.

Provides enterprise-grade security and compliance features including SOC 2 certification, HIPAA compliance, GDPR data handling, and audit logging. The platform supports role-based access control, data encryption at rest and in transit, and compliance reporting. Specific implementation details are not publicly documented but are available under NDA for enterprise customers.

Processes 22+ file formats including PDFs, images (PNG, JPEG), tables (XLSX), presentations (PPTX), and structured data (CSV, XML, JSON) into a unified searchable index. The system extracts text from images using OCR, parses table structures, preserves formatting metadata, and creates embeddings for both text and visual content. Retrieved results include the original visual elements alongside text, enabling questions about charts, diagrams, and images.

Enables filtering retrieved documents by custom metadata (key-value pairs) attached during ingestion, allowing queries like 'find documents from Q3 2024 with department=finance'. Metadata is indexed alongside embeddings, enabling combined semantic + metadata filtering in a single query. Supports boolean operators (AND, OR, NOT) and range queries on numeric metadata.

Maintains multi-turn conversation state where each user message is augmented with retrieved context from the knowledge base before being sent to the LLM. The system retrieves relevant documents for each turn, appends them to the conversation history, and passes the enriched context to the LLM for response generation. This enables coherent multi-turn Q&A where the LLM can reference both previous conversation turns and retrieved documents.

Integrates with external data sources (Google Drive, SharePoint, Notion) via pre-configured connectors that automatically crawl and ingest documents on a schedule. The system maintains a mapping between source documents and ingested chunks, enabling automatic updates when source documents change. Connectors handle authentication, pagination, and format conversion without requiring manual intervention.

Abstracts LLM provider selection, allowing users to configure different LLM backends (OpenAI, Anthropic Claude, Google AI, xAI Grok, Azure, Cohere, Qwen, Mistral, DeepSeek) without changing application code. The system handles provider-specific API differences, token counting, and response formatting transparently. Users specify model via configuration, and the platform routes requests to the appropriate provider.

Provides TypeScript and Python SDKs with native bindings to Vercel's AI SDK, enabling seamless integration into existing AI applications. The SDK abstracts HTTP calls to the Agentset API, handles authentication, manages request/response serialization, and provides type-safe interfaces (TypeScript). AI SDK integration enables use of Agentset as a tool within AI SDK agent frameworks.

Exposes Agentset as an MCP (Model Context Protocol) server, enabling external applications and LLM clients to query the knowledge base through a standardized protocol. The MCP server implements Agentset's search and retrieval capabilities as MCP tools, allowing any MCP-compatible client (Claude, other LLMs, custom agents) to access the knowledge base without direct API integration.

An enterprise-tier feature enabling extended multi-step reasoning over documents with configurable depth and breadth. The system performs iterative retrieval and synthesis with explicit reasoning steps, potentially including hypothesis generation, evidence gathering, and conclusion refinement. Specific implementation details are not publicly documented, but benchmarking on FinanceBench suggests capability for complex financial analysis.