Skip to main content
For the complete documentation index for agents and LLMs, see llms.txt.

Feature List

Haystack Enterprise Platform is an end-to-end platform for building, deploying, and managing AI applications in production. From RAG systems to AI agents, it gives you full control over every layer of the stack while handling the infrastructure complexity for you.

Haystack Enterprise Platform is built on Haystack, deepset's open source Python framework. Pipelines and agents you build in the platform use the same components and pipeline structure as Haystack. To learn more, see Haystack Enterprise Platform and Haystack.

You can run the platform in the managed cloud or self-host it on your own infrastructure, including on-premises and air-gapped environments.

AI Agents​

Build autonomous, multi-step AI systems that reason, plan, and act. The Haystack Platform Agent uses an LLM to decide which tools to call and loops until it reaches a result, handling tasks far beyond simple question answering. Our agents are:

  • Model-agnostic: Use any LLM as the Agent's brain and easily swap them out.
  • Rich tool ecosystem: Give your Agent access to pipelines, custom Python functions, and MCP servers as tools.
  • Controllable: Define exit conditions so the Agent stops exactly when and how you want.
  • Context-aware: Built-in conversation memory tracks history and tool usage across turns. You can also choose how much of the previous turn, including tool calls and reasoning steps, an Agent sees on each follow-up request. Beta
  • Long-term memory: Give an Agent Mem0 memory tools to store and retrieve facts and preferences across conversations, kept separate for each user. For details, see Agent Memory.
  • Real-time streaming: Stream Agent responses as they're generated.
  • Full observability: Trace every Agent decision with built-in traces, Langfuse, or Weights & Biases Weave.
  • Governed: Limit steps and tools, control who can deploy and run each Agent, and review every run. For details, see Agent Governance and Control.

Start with ready-made Agent templates and customize from there. For details, see AI Agents and Building AI Agents.

Visual Pipeline Builder​

Design AI applications visually with Pipeline Builder, an intuitive drag-and-drop editor.

  • Drag and drop from a library of 180+ pre-built components to assemble your pipeline.
  • Switch between visual editing and YAML with real-time synchronization.
  • Build complex flows with branching, loops, and conditional routing.
  • Export to YAML or Python for local development and version control.
  • Jump-start with curated, tested pipeline templates covering RAG, agents, data processing, and more.
  • Inspect and fix all validation and runtime issues in one place.
  • Find components with natural language search in the component library, or press Cmd+K to search the components in your pipeline and create connections.
  • Organize components in collapsible groups, and undo and redo your changes.
  • Remove glue components with smart connections: connect LLM or Agent directly to Input and Output. For details, see Simplify Pipelines.
  • Debug with the AI Assistant: describe a problem or click Fix with AI next to an issue, and the assistant suggests a fix and can apply it. For details, see Debug with Builder.
  • Work in light or dark theme, or follow your system setting.

To learn more, see Create a Pipeline in Studio. To understand pipelines, see Pipelines

Model Flexibility​

Avoid vendor lock-in. Haystack Enterprise Platform is model-agnostic, so you can swap LLMs, embedding and ranking models without changing your pipeline architecture.

Supported providers include OpenAI, Anthropic, Azure OpenAI, Amazon Bedrock, Google Gemini, Cohere, Mistral, NVIDIA, Meta Llama, Together AI, and more. New models are added regularly, including Claude 5, GPT-5.6, and Gemini 3.5 models, with parameters to control reasoning depth. Save custom model definitions with your preferred parameters, including models hosted through Anthropic on Azure, and reuse them across pipelines at the workspace or organization level. For details, see Add Custom Model Definitions.

For the full list of providers, see Supported Connections and Integrations.

Multimodal AI​

Build systems that go beyond text. Process, understand, and generate content across multiple data types, including audio, images, and documents:

  • Audio: Transcribe speech to text with components like Whisper and build search over audio content.
  • Images: Use multimodal LLMs (GPT-5.5, Claude, Gemini, and others) to analyze and reason over images.
  • Documents: Extract structured data from PDFs, Office files, and scanned documents using Azure Document Intelligence, Unstructured.io, or Docling.

For more details, see Multimodal Systems.

MCP Support​

Connect your AI applications to external systems through the Model Context Protocol (MCP). Use the official deepset MCP server to manage Haystack Platform resources from your coding agent, or connect your agents to third-party MCP servers.

You can also use your deployed pipelines as MCP tools in AI coding assistants such as Cursor, VS Code, and Claude Code. Create one MCP server for your workspace, turn on each pipeline you want to expose, and give each tool its own name, description, and instructions. For details, see Use Your Pipelines as MCP Tools.

For more information, see Agent Tools and Model Context Protocol.

Data Ingestion and Processing​

Bring in data from any source and prepare it for your AI applications:

  • File processing: Handle PDFs, Office documents, HTML, images, and audio with out-of-the-box converters and preprocessors. For details, see Preparing Your Data.
  • External services: Process files with Azure Document Intelligence, Unstructured.io, or DeepL for translation.
  • Metadata: Attach, filter, and use metadata throughout your pipelines. Add metadata when you upload files and download it from the Files page. For details, see Add Metadata to Your Files.
  • Classification and extraction: Detect languages, classify documents, and extract named entities with language detection, zero-shot classification, and named entity extraction components.
  • More components: Use Apache Tika and Azure OCR converters, web search components such as Serper and SearchApi, and OpenAPI connectors. For the full list, see Integrations Overview.

Flexible Document Stores​

Choose where your data lives. Haystack Enterprise Platform supports multiple document stores so you can use the vector database that fits your stack:

  • Managed: OpenSearch (fully managed by Haystack Platform).
  • Bring your own: Elasticsearch, MongoDB Atlas, Pinecone, Qdrant, Weaviate, PGVector, ArangoDB.
  • Multiple sources: Query several indexes and document stores in one pipeline step with MultiRetriever. Mix keyword, embedding, and hybrid retrieval across sources, and choose how results are merged. For details, see MultiRetriever.
  • Databases: Connect to Snowflake for structured data queries.

You can also bring your own S3 bucket or OpenSearch cluster through VPC integration. For an overview, see Document Stores.

Custom Components​

Extend the platform with your own logic. Write custom Python components and use them alongside built-in ones in your pipelines.

  • Create reusable, versioned components shared across your organization.
  • Save a Code component as a reusable custom component for your workspace or organization, and set its init parameters from the UI.
  • Use the inline Code component for quick, single-pipeline customizations.
  • Securely manage API keys with secrets—no credentials in code.

For details, see Add Custom Code.

Prompt Engineering​

Iterate on prompts faster with built-in tools:

  • Test and refine prompts in Prompt Explorer with live pipeline responses.
  • Compare prompt performance side by side across up to three pipelines.
  • Save prompts to a library and update them across pipelines without redeployment.
  • Save a prompt update as a new pipeline version, so the deployed version stays unchanged until you deploy the new one.

For details, see Engineering Prompts.

Production-Grade Infrastructure​

Deploy with confidence. Haystack Enterprise Platform handles scaling, reliability, and performance so you can focus on your application.

  • One-click deployment: Move from prototype to production in seconds.
  • Autoscaling: Pipelines scale automatically with traffic (up to ten replicas for production workloads). You can configure replicas and standby behavior in the pipeline's Settings tab.
  • Service levels: Mark pipelines as Development or Production to control standby behavior and resource allocation. For details, see Pipeline Service Levels.
  • Pipeline versions: Save drafts, finalize them as numbered versions, compare and restore versions, and set the default version that endpoints use when a request doesn't name one. Test any version in the Playground. For details, see Manage Pipeline Versions.
  • Haystack version and dependencies: Pin the Haystack version a pipeline runs on, and add approved Python packages with the dependencies block in YAML. The platform runs the pipeline in an isolated environment and checks the block when you save or deploy. For details, see Update Pipeline Settings.
  • Streaming: Stream responses from pipelines and choose which components stream. Streaming events tell your client when a new replica is still warming up. For details, see Enable Streaming.
  • CI/CD: Automate pipeline deployment with GitHub Actions.

Deployment Options​

Run the platform where your data and compliance requirements need it to run:

  • Managed cloud: deepset operates and updates the platform for you.
  • Self-hosted and on-premises: Run the platform in your private cloud or data center.
  • Air-gapped: Run the platform with no internet connection, using models you host inside your network.

For a comparison of the options and how they affect data residency, see Deployment Options.

Observability and Tracing​

Understand exactly what happens inside your AI applications:

  • Built-in traces: Browse every execution of a pipeline in the Traces tab, filter by status and duration, and open a trace to see its span tree, waterfall timeline, logs, inputs, outputs, and failures. Export a trace as JSON, or jump between a trace and its search history entry. For details, see Trace with Built-In Traces.
  • Pipeline logs: View component-level logs with timestamps, messages, and log levels.
  • Live debugger: Run your pipeline step by step in Pipeline Builder, set breakpoints on any component, review the captured snapshot, change its inputs, and resume. Attach files to debug runs for pipelines with file inputs.
  • Search history: Review every query with its status, answer, rating, and feedback. Add notes and hashtags, choose the columns you see, and export to CSV with the pipeline version that served each query. For details, see Monitor Pipeline Performance.
  • OpenTelemetry and Datadog: Send pipeline telemetry to your own monitoring stack with the OpenTelemetryConnector and DatadogConnector components.
  • Langfuse integration: Trace full query journeys with spans, latency breakdowns, and dependency maps. For details, see Trace with Langfuse.
  • Weights & Biases Weave: Monitor ML telemetry, token counts, and performance metrics. For details, see Use Weights & Biases.
  • Remote debugging: Connect to running pipelines through a VS Code tunnel.

For more information on observability, see Trace Your Pipelines.

Enterprise Integrations​

Connect SharePoint, Google Drive, HTTP APIs, MCP, and sync patterns. For details, see Integrate with Enterprise Systems.

User Feedback​

Test your AI applications with real users and improve continuously:

  • Shareable prototypes: Generate branded, shareable links to your pipeline—no login required for testers. Customize with your logo, in JPEG, PNG, or SVG format, and your colors. For more information, see Share a Pipeline Prototype.
  • Structured feedback: Collect ratings, tags, and comments from users. Group, filter, and export feedback for analysis. For details, check Collect User Feedback.
  • Feedback insights: Analyze feedback patterns and generate summaries with AI-powered Feedback Insights in the search history.
  • Playground: Test pipeline responses interactively before sharing, choose the pipeline version to test, preview uploaded images and files, and download an answer as a Markdown file. For instructions how to do it, see Testing Your Pipeline.

Batch Processing with Jobs​

Run queries at scale using Jobs:

  • Process query sets in bulk across your entire dataset.
  • Run queries once on all files or repeat per individual file.
  • Share formatted results with anyone—no login required.
  • Organization admins can see the credits each job used on its Job Details page.

Enterprise Security and Access Control​

Keep your data and applications secure with enterprise-grade access management:

  • Role-based access control: Assign preset or custom roles with granular permissions at the organization and workspace levels. Editors can manage workspace members and workspace custom models. For details, see User Roles and Permissions and Enterprise Security and Governance.
  • Secrets management: Store API keys and credentials securely, scoped to workspaces or organizations. To learn about secrets, see Secrets and Integrations.
  • VPC integration: Run with your own OpenSearch cluster and S3 bucket for full data isolation. For details, see Connect Your Own File Storage.
  • Workspace isolation: Organize teams and projects into isolated workspaces (up to 100 per organization).
  • Agent governance: Set step limits, exit conditions, and tool lists for each agent, and map them to governance requirements. For details, see Agent Governance and Control.
  • Certifications: deepset is SOC 2 Type II and ISO 27001 certified, and the platform is GDPR and HIPAA compliant.

REST API and SDK​

Integrate Haystack Enterprise Platform into your applications and workflows:

  • REST API: Full programmatic access to pipelines, files, indexes, jobs, and feedback. For full reference, see API.
  • Chat and streaming endpoints: Create search sessions automatically, continue a conversation with a session ID, and turn off keep-alive comments for clients that only read data lines.
  • Python SDK: Deploy and run pipelines, upload files, and automate workflows with Python or the CLI. To learn more, see Haystack Enterprise SDK.
  • MCP server: Expose Haystack Platform pipelines as tools for any MCP-compatible agent.

For licensing information about third-party software used in Haystack Enterprise Platform, see Third Party Software.