Haystack is coming to @NVIDIAGTC Berlin 💚💙

🌃 Tomorrow, 21 Oct, 18:00: Hack Night at our Berlin office, part of our House of Sovereignty series. Get hands-on with the Haystack Enterprise Platform, take on a challenge, and meet the team. Only 30 spots, so apply here:
→ luma.com/hep-hack-night

🎙️ Thursday, 22 Oct, 10:00: @malte_pietsch and @Julian_Risch join Oleg Sudakov from @nvidia for "Evolve Within Bounds: Secure, Self-Improving Agents for On-Prem AI"
→ nvidia.com/en-eu/gtc/sess…

📍 Booth 3071, both expo days: come meet the team and grab your Hay-swag

Haystack 3.4 is here, and it focuses on agents that can run for long periods.

Agents that call search, file, or API tools can quickly fill the context with large results. 3.4 adds the experimental ToolResultOffloadCompactor for CompactionHook. When compaction triggers, it writes store and replaces them with references. The agent reads them back only when it needs them. Context is shorter, but nothing is lost.

Other highlights:
🔍 Better observability: failed tool calls carry a haystack.tool.error tag in tracing spans, so they’re no longer confused with tools that return error dicts
🐍 Python 3.15 support (Python 3.10 is no longer supported because it reaches EOL this month)
📄 Converter improvements: DOCX text boxes are extracted, HTML encodings are detected automatically

⚠️ Before you upgrade ⚠️
The default OpenAI and Azure embedding model is now text-embedding-3-small. It’s cheaper and stronger, but its embeddings aren’t compatible with ada-002. If you have an existing index, pin the old model or re-embed.

💻 pip install -U haystack-ai

Full release notes: haystack.deepset.ai/release-notes/…

Most of what we ask LLMs to do inside pipelines isn't generation. It's decisions. @typesafeai System One models are built for exactly that, and they're now in @Haystack_AI

The new integration gives you two components:

🏷️ TypeSafeDocumentClassifier answers your questions about s the results in its metadata.
🔀 TypeSafeTextRouter sends text to the right branch, with an optional low_confidence output so uncertain cases can fall back to an LLM or a human.

Use them with TypeSafe's hosted Jev model, or keep everything on your own infrastructure with Ollaya, which serves open decision models like Laya locally through the TypeSafe API.

🐍 pip install typesafe-haystack
👉 haystack.deepset.ai/integrations/t…

Europe has a new open-weight model for sovereign AI 🇪🇺

@Aleph__Alpha's Kolibri works with @Haystack_AI from day one.

Developed and trained in Europe, Kolibri is a German-English Mixture-of-Experts model released under Apache 2.0 and built for demanding use cases across public stry.

With Haystack, you can run Kolibri on infrastructure you control and build it directly into your own Agents and pipelines, without making an external model API a dependency.

Open framework. Open weights. Your infrastructure. Your AI stack.
haystack.deepset.ai/integrations/k…

🌐 Building an AI agent that needs web intelligence?

@youdotcom brings semantic web search and real-time retrieval directly into your Haystack agents!

Add YoucomWebSearch to your agent (or your pipeline), configure your You.com API credentials, and your agent mid-run, pulling in content, snippets, and metadata that provide enough ground for downstream tasks.

Think of an agent that researches a topic in real time instead of working off its index.

🐍 pip install youcom-haystack
🔗 haystack.deepset.ai/integrations/y…

Haystack 3.3.0 is here 🚀

This one's a focused release, with a nice performance win at the center:

⚡️ Faster SentenceWindowRetriever
It now queries the Document Store once per run call instead of once per retrieved document. Fewer calls = lower latency & less load, and your pipege, so it's a free speed-up on upgrade.

Plus more precise BM25 ranking, enhancements, security patches, and bug fixes!

💙 Big thanks to our contributors to this release!
↓ Full release notes 🧵

📊 @rhesisai is an open-source platform for collaboratively testing and evaluating LLM applications and agents.

It is now available in @Haystack_AI via RhesisConnector. Constructing the connector installs the tracer, so every pipeline run, agent step, tool call, and specialist is as a single trace, grouped by session.

The connector covers both sides of a release. During testing, traces show why a case failed and where in the pipeline it failed. In production, they show the same for real user sessions, and a session that went wrong becomes the next test case.

🐍 pip install rhesis-haystack
🔗 haystack.deepset.ai/integrations/r…

Haystack 3.2 is here 🚀

This release keeps agents lean and in control as conversations grow. Three highlights:

🧵 SummarizationCompactor for smarter context compaction
Pair SummarizationCompactor with CompactionHook (shipped in v3.1) to progressively summarize long-running TokenBudgetHook caps what a run can spend
A ready-made hook that ends the run at "token_budget_exceeded" once you hit your token ceiling through the stop_run hook point.

🛠️ Faster Pipeline building
New add_components() and connect_many() let you add and wire up multiple components in one call.

💙 70 contributors made this release happen; huge thanks to everyone!

🔗 Full release notes: haystack.deepset.ai/release-notes/…

Full-text search without semantic understanding leaves gaps. Dense vector search without keyword matching misses precision.

@ApacheSolr is now available as a Document Store and retriever in Haystack. Use SolrDocumentStore with SolrHybridRetriever to combine BM25 full-text search e pipeline - getting both lexical and semantic matching without extra infrastructure.

Solr has been a trusted open-source search engine for over a decade, and its hybrid retrieval approach fits naturally into Haystack pipelines whether you're prototyping locally or scaling to production.

🐍 pip install solr-haystack
🔗 haystack.deepset.ai/integrations/s…

📣 We're hosting our next unconference in Berlin with @prior_labs, exploring open-weight models and open-source agents! Come join us if you'd like to discuss:

→ Structured data in agentic workflows
→ When agents should hand off to TabPFN for prediction
→ Benchmarking tabular rchestration matters for production agents
→ What owning your stack end-to-end actually requires
→ Open-weight vs. closed models in practice

Or anything else that's been challenging you lately!

📍 @deepset_ai HQ, Berlin
🗓️ September 24th, 6PM-9PM

Register: luma.com/haystack-prior…

If you are a @mariadb user, you can now store your embeddings right alongside your documents and metadata in the same system - no need for a separate vector database. MariaDBDocumentStore and MariaDBEmbeddingRetriever let you build semantic search pipelines that keep all your o manage.

Embed your content, persist vectors in the same MariaDB instance where your source data lives, and retrieve similar results in seconds. Everything stays in one place.

🐍 pip install mariadb-haystack
🔗 haystack.deepset.ai/integrations/m…

Haystack 3.1 is here 🚀

This release is all about smarter context management with token accounting, context compaction & precision. Three highlights:

🪝 Context compaction for agents
New CompactionHook automatically shortens long-running conversations before they blow through the token counters for estimating usage
Three new counters, from a zero-dependency approximation to an exact, OpenAI API-backed count, so you know when to compact before you hit the wall.

🛠️ AgentTool for multi-agent systems
Wrap any Agent as a Tool so another Agent can delegate to it. Only the final reply crosses back, the coordinator's context stays clean.

💙 42 contributors made this release happen; huge thanks to everyone!

👉 Full release notes: haystack.deepset.ai/release-notes/…

HetznerChatGenerator brings European sovereign cloud LLMs to Haystack. Process text and images through @Hetzner_Online infrastructure while keeping your AI workloads within EU borders.

🐍 pip install hetzner-haystack
🔗 haystack.deepset.ai/integrations/h…

We are thrilled to be among the ten finalists for The Spark – The Technology Award.

For eleven years, @handelsblatt and @McKinsey have been recognizing Germany’s most innovative technology startups.

This year’s theme is "Scaling Deep Tech", focusing on technologies that are orld impact today and shaping the future of industry and the economy.

Now, we need your support.

Vote by September 7th to help us reach the Top 3! Your vote counts directly toward the jury's decision.

Vote for us here: cmk.handelsblatt.com/cms/articles/2…

Finally, @Linkup_platform joins Haystack's growing search component library. Fetch real-time web results, get ranked documents with source links, and integrate them into your pipeline in a few lines. Same philosophy as other Haystack components - simple, direct, pluggable.
haystack.deepset.ai/integrations/l…

📷 DDGS (Dux Distributed Global Search) is now available in Haystack. Query Google, Bing, Brave, Yahoo, Yandex and more from your pipelines and agents. Multi-source web search without the infrastructure overhead.

pip install ddgs-haystack
haystack.deepset.ai/integrations/d…

Please welcome Mirage by @struktoai to the Haystack integrations family. Create isolated virtual filesystems with multiple mounts, such as in-memory storage backed by cloud resources like S3, and give your agents the ability to explore and analyze files using bash commands nfrastructure.

Agents can now run grep, cat, head, wc, and other shell utilities against mounted data sources, treating them as a unified filesystem. Control which commands are allowed and keep sensitive resources read-only while giving agents just the access they need.

🐍 pip install mirage-haystack
🔗 haystack.deepset.ai/integrations/m…

Haystack 3.0 shipped this month. We spent launch week showing you what it can do. Now here's what it took to build it 👇

Since @Haystack_AI 2.0: 2,400+ pull requests merged, 195 contributors, ~170 of them making their first-ever contribution along the way. 17,258 CI runs since roughly one every five minutes, around the clock, for eight weeks.

Most-edited file of the whole cycle: agent.py. Of course it was.

Migrating to 3.0? The full guide covers all breaking changes: docs.haystack.deepset.ai/docs/migration
Or skip straight to haystack-v2-to-v3, an agent skill that migrates your v2 pipelines and agents for you.

💙 Huge thanks to our open-source team at @deepset_ai, and to our community contributors who shipped fixes and features for v3 alongside us! Haystack 3.0 is yours as much as it's ours.

Got questions about migrating, or curious why we made a specific call? Join us at Haystack 3.0 Office Hours next week, August 4, with @LukawskiKacper and @bilgeycl
👉🏼 Register: luma.com/haystack-3

Day 5 of Haystack 3.0 Launch Week: Human-in-the-Loop from Terminal to Production

For today's drop, we implemented a production agent that pauses before anything risky and waits for a human to say go. The approval shows up right where the conversation is happening, making the he users. We built the whole thing as a real, deployable service (Hayhooks + @Redisinc + @OpenWebUI), so it's not just another notebook trick or a simple terminal demo

Set up with one command and run all services with @Docker 🐳

🎥 Watch it in action ↓

Deep Research Agent

Advanced RAG Agent

Agent Pack ships new with @Haystack_AI 3.0: a collection of complex, pre-configured Haystack agents you can run as they are, customize, or copy as a blueprint for your own architecture.

📬 On Day 3/5 of Launch Week, we're highlighting these pre-built agents:
- Advanced RAG RAG agent that constructs its own filters to narrow retrieval
- Deep Research Agent: a multi-agent system that writes a cited Markdown report, using @tavilyai under the hood for search

Two agents today, more landing in the pack over time.
Let's see how they work under the hood 🧵

Your agent doesn't check with you before it makes its next move. It just runs.

That's the agent loop: LLM call, tool call, LLM call, tool call, however many times it takes to finish. Most frameworks let you watch it happen. Very few let you step into it.

Day 2 of Haystack 3.0 ntrolling the agent loop for token cost, not just observing it.

Every `Agent.run()` already returns `step_count`, `token_usage`, and `tool_call_counts`. Agent hooks let you act on that metadata from inside the loop itself:
→ Soft limit warns, hard limit blocks
→ `before_llm` stops an over-budget run before its next LLM call
→ `after_run` gives you a final report every time, clean, warned, or stopped mid-loop

🎥 Watch it in action ↓

Haystack 3.0 is here 🚀

Agents move to the center of the framework:
🤖 Pre-built agents: deep research + advanced RAG, ready out of the box
🪝 Hooks to control the agent loop
🧰 First-class skills with progressive disclosure
📉 A leaner core: we deleted 3 lines for every 1 we Day 1. New drops every day this week at 3 PM CEST.
Full announcement 👇
haystack.deepset.ai/blog/haystack-…

3 days until Haystack 3.0 🚀
Monday, 3 PM CEST | 9AM EST. Then a new drop every day, all week.

The countdown is running ⏳

Integration alert! @OrcaRouter is now available as a ChatGenerator in Haystack. Use OrcaRouterChatGenerator to swap between standard LLMs, and multimodal models through a unified interface - no pipeline rewiring needed.

It acts as a flexible proxy layer that handles tool and multimodal inputs seamlessly, letting you experiment with different models or providers without changing your agent code.

🐍 pip install orcarouter-haystack
🔗 haystack.deepset.ai/integrations/o…

Your @googledrive is now queryable in Haystack. Use GoogleDriveRetriever to search files stored in Google Drive, then GoogleDriveFetcher to retrieve their content. Built on the OAuth integration, it handles secure token management automatically so your pipelines stay interact with your application.

This integration is ideal for RAG applications that need to pull knowledge from shared team documents, spreadsheets, or presentations stored in Google Drive without duplicating data into a separate vector database.

🐍 pip install google-drive-haystack
🔗 haystack.deepset.ai/integrations/g…

Big news: Haystack 3.0 ships Monday, July 20 🚀

It's one of the biggest releases in Haystack's history, and one announcement wouldn't do it justice. So we're running our first-ever Launch Week: July 20 to 24, five days, five drops, something new every day at 3 PM CEST 🎉

Day 1 he release itself and the full story of what changed and why. After that, each day goes deeper into a different part of 3.0. We won't spoil the lineup, but if you've been following Haystack on GitHub recently, you might see some of it coming 👀

Check the countdown and sign up for updates 👇
haystack.deepset.ai/launch-week/ha…

See you Monday 💙

Search and fetch @SharePoint documents directly in Haystack with the microsoft-sharepoint-haystack integration. MSSharePointRetriever queries across your SharePoint sites while MSSharePointFetcher downloads full document content, both powered by the OAuth integration for ent. Perfect for RAG pipelines that need to tap into enterprise document libraries stored in @Microsoft365 without friction.

🐍 pip install microsoft-sharepoint-haystack
🔗 haystack.deepset.ai/integrations/m…

Secure OAuth flows are now built into Haystack. The OAuthTokenResolver handles token exchange automatically - connect to Microsoft Graph, Google APIs, and any OAuth provider without manual credential management.

OAuth is the standard. It powers authentication across the web. Now can leverage it natively.

🐍 pip install oauth-haystack
🔗 haystack.deepset.ai/integrations/o…

Haystack 2.31.0 is here 🚀

This is the final 2.x release before Haystack 3.0. It focuses on preparing the ecosystem for the next major version while delivering quality-of-life improvements today. Many interesting updates this time, but only two highlights.

🧠 @cognee_ is now available as a memory store integration in Haystack. Seed permanent knowledge into your graph and then retrieve relevant memories in agentic pipelines. With Cognee and Haystack, you can:

- Write domain knowledge once to the permanent knowledge graph
- Build versation history with long-term facts
- Create session-scoped stores with just a session_id string

This makes it simple to build stateful agents that remember facts across sessions and grow smarter as they interact, combining real-time conversation with persistent domain knowledge in a single unified system.

🐍 pip install cognee-haystack
🔗 haystack.deepset.ai/integrations/c…

Another integration alert! Use FunASRTranscriber to convert audio files to Documents with fully self-hosted speech recognition supporting 50+ languages. No API keys required.

FunASR handles transcription, speaker diarization, and timestamp extraction. Pair it with Haystack o-to-document workflows that stay entirely on your infrastructure. Models are downloaded on first use and cached locally.

@ModelScope2022

🐍 pip install funasr-haystack
🔗 haystack.deepset.ai/integrations/f…

ArangoDB is available as a Document Store in Haystack. Use ArangoDocumentStore to connect directly to your ArangoDB instance, and integrate it into RAG pipelines that leverage ArangoDB's multi-model capabilities.

@arangoai flexible schema and native support for documents, e it ideal for applications that need both structured data storage and semantic search. Whether you're indexing documents, managing relational data with embeddings, or building hybrid search experiences, ArangoDB handles it seamlessly.

🐍 pip install arangodb-haystack
🔗 haystack.deepset.ai/integrations/a…

A new Document Store has been added to Haystack. Use @supabase to build pgvector vector search capabilities directly into your Haystack pipelines, use Groonga for efficient full-text search, or choose Supabase Storage for seamless file handling.

@supabase makes it easy to host the cloud without managing infrastructure - just spin up a project, get connection credentials, and you're ready to embed and search. This works well for applications needing scalable vector storage with the familiarity and power of @PostgreSQL

🐍 pip install supabase-haystack
🔗 haystack.deepset.ai/integrations/s…

Extract text, tables, and forms from images and PDFs with Amazon Textract in Haystack. AmazonTextractConverter brings @awscloud's OCR capabilities directly into your document processing pipelines. Just pass an image or single-page PDF and ask natural-language questions to extract ically.

🐍 pip install amazon-textract-haystack 🔗 haystack.deepset.ai/integrations/a…

👋 @LiteLLM is now available as a Chat Generator in Haystack. Access 100+ LLM providers from a single interface - OpenAI, Anthropic, Groq, Cohere, and more. Switch models and providers without rewriting code.

🐍 pip install litellm-haystack
🔗 haystack.deepset.ai/integrations/l…

Haystack now publishes a public Model Context Protocol server. Point Claude Code, Cursor, or any MCP-compatible coding agent at it and get real-time access to the latest Haystack docs - no API keys or sign-ups needed.

🖇️ docs.haystack.deepset.ai/docs/docs-mcp-…

Haystack 2.30 is here 📷

Smarter code processing and simpler APIs. Two highlights this time.

Vespa is now available as a Document Store in Haystack. Use VespaDocumentStore for hybrid and semantic search with a powerful, production-ready engine, and pair it with VespaEmbeddingRetriever to index and retrieve documents directly in your pipelines. Metadata filtering excels at large-scale information retrieval with advanced features like real-time indexing, multi-modal search, and distributed document management - ideal for applications that demand both speed and sophistication.

🐍 pip install vespa-haystack
🔗 haystack.deepset.ai/integrations/v…

We're launching the Haystack Ambassador Program 🎉

It's for the people already building, teaching, contributing to the Haystack ecosystem and who want to go further.

🧵 Here's what it is, who it's for, and how to apply

🌐 @perplexity_ai is now integrated with @Haystack_AI!

- `PerplexityWebSearch` to fetch fresh web results directly into your pipelines
- `PerplexityChatGenerator` to power agent conversations with Perplexity's OpenAI-compatible LLM API
- `PerplexityTextEmbedder` and dder` for semantic embeddings

The web search component is especially useful when your agents need real-time information beyond their training data. Just instantiate it with your API key, set `top_k` results, and pass your query. You get back both documents and source links in a single call. Perplexity's API combines language models with live web access, making it a strong choice for building knowledge-intensive applications that need current information.

🐍 pip install perplexity-haystack
🧩 haystack.deepset.ai/integrations/p…
📖 docs.perplexity.ai

We're so happy that @e2b is now available as a tool integration in Haystack. Give your agents direct access to isolated Linux sandboxes for code execution, file operations, and command-line interactions - all without leaving your pipeline.

E2BToolset lets agents write scripts, ilesystems inside secure, ephemeral environments. Perfect for coding assistants, data processing tasks, or any agent that needs to execute code safely.

🐍 pip install e2b-haystack
🔗 haystack.deepset.ai/integrations/e…

Use AlloyDBDocumentStore to build RAG applications backed by Google Cloud's PostgreSQL-compatible database, with native vector search and full Haystack pipeline integration. Perfect for enterprises running workloads on @googlecloud who need scalable, managed vector storage ta.

🐍 pip install alloydb-haystack
🔗 Documentation: haystack.deepset.ai/integrations/a…

DoclingServeConverter integrates docling-serve into Haystack pipelines via HTTP, eliminating the need to embed Docling's dependencies in your application. Configure the converter with a docling-serve endpoint, send documents through your pipeline, and receive structured output re information. This architecture decouples document processing from your main service, enabling independent scaling and reducing deployment overhead.

🐍 pip install docling-serve-haystack
🔗 haystack.deepset.ai/integrations/d…

@LFAIDataFdn

FalkorDB is now available as a Document Store in Haystack. Use FalkorDBDocumentStore to store documents and vectors in a graph format, pair it with FalkorDBEmbeddingRetriever to retrieve documents by semantic similarity, and use FalkorDBCypherRetriever for graph-based queries.
build retrieval pipelines that leverage both vector search and graph relationships. @falkordb is built as an extension of Redis, making it lightweight and efficient for applications that need real-time graph and vector capabilities.

🐍 pip install falkordb-haystack 🔗 haystack.deepset.ai/integrations/f…

@brave Search is now available as a web search component in Haystack. Use BraveWebSearch to retrieve relevant documents from the web and pipe them directly into your LLM-powered applications. It integrates seamlessly with Haystack's Pipeline, making it easy to build augmented at leverage real-time web search results.

@brave Search combines privacy-respecting search with developer-friendly API design. It's ideal for building agents, question-answering systems, or any application that needs fresh information from the web without relying on a separate search service.

🐍 pip install brave-search-haystack
🔗 haystack.deepset.ai/integrations/b…

Asqav is now available as a governance and audit trail integration for Haystack. Use AsqavComponent to sign data flowing through your pipelines with ML-DSA-65 post-quantum signatures and create tamper-evident audit records. Each signed action returns a signature_id linking to its ides governance and compliance capabilities for organizations that need auditable, tamper-proof records of data processing in their AI systems.

🐍 pip install asqav
🔗 haystack.deepset.ai/integrations/a…

@ChonkieAI is now available as a text-splitting integration in Haystack. Use ChonkieTokenDocumentSplitter and other chunkers from Chonkie's library to intelligently segment your documents before indexing. The integration includes RecursiveChunker, SemanticChunker, and additional match your workflow.

@ChonkieAI's chunkers are designed to preserve semantic boundaries and handle various text types. Whether you need simple token-based splitting or sophisticated semantic awareness, the integration exposes multiple chunking strategies that can be wired into your preprocessing pipelines and retrieval systems.

@ChonkieAI has become a popular choice in RAG systems for its ability to balance chunk quality with performance, and now it integrates seamlessly into Haystack's document processing workflows.

🐍 pip install chonkie-haystack 🔗 haystack.deepset.ai/integrations/c…

Haystack 2.29 is here 🚀

This release brings hybrid search and flexible LLM messaging to the core, plus async support for cache checking. Many interesting updates this time, but only one highlight:

🔍 Combine Retrievers with MultiRetriever Run multiple text retrievers in r results into a deduplicated list ranked by reciprocal rank fusion. Toggle individual retrievers at runtime using the active_retrievers parameter - skip the embedding retriever for keyword-only queries, for example.

💙 Big thanks to our contributors to this release!

👇 Full release notes below

Presidio is now available as a PII detection and anonymization integration in Haystack. Use PresidioEntityExtractor to identify sensitive information like email addresses, phone numbers, names, and credit card numbers in your documents and text. Pair it with or PresidioTextCleaner to automatically redact or mask that data before it flows through your pipelines.

This is essential for applications handling user data at scale - financial systems, healthcare platforms, customer support tools, or any RAG pipeline that processes documents with personal information. Configure entity types and confidence thresholds to tune detection precision for your use case.

Presidio, built and maintained by @Microsoft, has become the industry standard for PII detection and is now deeply integrated into Haystack workflows. Multi-language support means your data governance works across global datasets.

⭐ @Haystack_AI just crossed 25,000 stars on GitHub!

This number means a lot to us, but what it really represents is you.

Every contributor who opened a pull request. Every community member who answered a question on Discord. Every developer who filed an issue, wrote a or simply built something incredible with Haystack and shared it with the world.

It's a true community effort 💙

When we first started Haystack, we believed that the journey to building great AI applications should be open, composable, and community-driven. 25,000 stars later, that belief has never felt more validated.

Thank you to every single one of you: contributors, users, advocates, and builders. Let’s keep going! 🚀

Haystack pipelines need strong retrieval primitives: embedders, rerankers, extractors. Today we're announcing the native Haystack and SIE integration from @superlinked: drop-in Haystack components for every stage of the pipeline, self-hosted on 85+ open-source models.

⚡ Thunderbolt is trending on GitHub today - and we're proud to be @mozilla launch partner.

MZLA (Mozilla's subsidiary behind Thunderbird) just launched Thunderbolt: an open-source, self-hostable AI client for enterprises and public sector organizations that need sovereignty tive apps across web, desktop, and mobile.

Haystack 2.28 is here 🚀

This release makes agent tool development more flexible and document processing more precise. Many interesting updates this time, but only one highlight:

🔗 Pass Agent State Directly to Tools

No more manual wiring. Tools and components can now declare eceive the live agent State at runtime automatically — giving them full access to conversation history and context without extra connections.

💙 Big thanks to our contributors to this release!

👇 Full release notes in the next post

Bigger context windows don't mean better agents - they mean more competition for the model's attention.

New post: context engineering for agentic systems by @LukawskiKacper. What fills the context window, why bloat hurts quality and cost, and how Haystack gives you full control - the infrastructure layer that keeps it all under control.

haystack.deepset.ai/blog/context-e…

Part 1 of a series - what would you most want us to cover next?

New tutorial: compress your LLM's KV cache with TurboQuant + Haystack 🗜️ by @LukawskiKacper

We used turboquant-vllm (a community TurboQuant implementation, based on the @GoogleResearch recent paper) hooked into HuggingFaceLocalChatGenerator, but since Haystack is hing compatible with HuggingFace, Ollama, or vLLM works just as well.

👉 haystack.deepset.ai/tutorials/49_t…

MarkItDown by @Microsoft quickly became one of the most talked-about document conversion libraries in the Python ecosystem. The reason is simple: it just works. PDF, DOCX, PPTX, XLSX, HTML, images - all converted to clean Markdown, locally, with no external API calls.

For RAG more than people realize. The quality of your conversions directly shapes the quality of your retrieval. Use MarkItDownConverter to convert virtually any file format into Haystack Document objects with Markdown content - standalone or wired directly into your indexing pipelines. Everything runs locally, so your data stays yours.

🐍 pip install markitdown-haystack
🔗 Documentation: haystack.deepset.ai/integrations/m…

Most document parsing pipelines still rely on cloud APIs, external services, or brittle format-specific libraries stitched together. @kreuzberg_dev takes a different approach: a Rust-core document intelligence engine that extracts text, tables, and metadata from 91+ file formats I calls. No data leaving your infrastructure.

We've now integrated it into Haystack as a converter component. Drop in KreuzbergConverter to transform PDFs, DOCX, PPTX, scanned images, emails, archives, notebooks, and more into Haystack Document objects.

🐍 pip install kreuzberg-haystack
🔗 Documentation: haystack.deepset.ai/integrations/k…

PII in RAG pipelines is one of those problems that's easy to ignore. A name, an email, a social security number slipping through into a vector store or an LLM prompt is the kind of thing that keeps compliance teams up at night. @tonicfakedata Textual has been tackling PII ation with transformer-based NER models that cover 46+ entity types across 50+ languages

We've now integrated it into Haystack with two components. TonicTextualDocumentCleaner sanitizes documents before they ever reach your index - swapping real PII with realistic synthetic data or reversible placeholder tokens. TonicTextualEntityExtractor detects PII and attaches it as structured metadata directly onto your documents, ready for hybrid retrieval, auditing, or compliance workflows.

🐍 pip install textual-haystack
🔗 Documentation: haystack.deepset.ai/integrations/t…

When your pipelines need to handle the messy reality of real-world documents like legacy .doc files, .ppt presentations, .ods spreadsheets, the LibreOfficeFileConverter takes care of the heavy lifting. Convert between formats, get back ByteStream objects, and feed them straight ipeline. Both sync and async execution modes are supported.

A practical example: convert a legacy .doc to .docx, then pass it directly to DOCXToDocument to extract clean Document objects - all in a single pipeline. No intermediate files, no friction.

🐍 pip install libreoffice-haystack

🔗 Documentation: haystack.deepset.ai/integrations/l…

TabPFN, developed by the @prior_labs team, brings the same "pretrain once, predict anywhere" idea that transformed NLP and vision, but applies it to the structured datasets that actually dominate enterprise ML. No feature engineering loops, no hyperparameter tuning, just upload ctions.

🐍 pip install mcp-haystack
🔗 Documentation: haystack.deepset.ai/integrations/p…

Haystack 2.27 is here 🚀

This release brings a couple of improvements that make building pipelines a lot smoother. Some highlights:

🔌 Automatic List Joining in Pipelines
No more adapter components just to combine inputs. Pipelines now automatically join multiple inputs into a in different but compatible types.

🗄️ Better Developer Experience for Document Stores
The metadata inspection and filtering utilities are now available in the InMemoryDocumentStore, enabling you to easily debug, filter, and inspect the data stored in the document store

💙 Big thanks to our contributors to this release!

👇 Full release notes in the next post

@arcade_db is a multi-model database that handles document storage, HNSW vector search, and SQL-based metadata filtering natively. It runs over a pure HTTP/JSON API.

We've now integrated it into Haystack as a Document Store. Use ArcadeDBDocumentStore to store and index your 's native HNSW vector index, and pair it with ArcadeDBEmbeddingRetriever to run similarity search with metadata filtering directly in your pipelines. Cosine, euclidean, and dot product similarity all supported out of the box.

🐍 pip install arcadedb-haystack
🔗 Documentation: haystack.deepset.ai/integrations/a…

🚀 Haystack 2.26 is out!

Here are a few highlights:

🤖 More flexible Agents with dynamic system prompts
You can now use Jinja2 templating in `system_prompt` to inject runtime variables and dynamically adapt behavior, such as changing language or tone, or adding conditional ng prompts.

🔍 LLMRanker for smarter reranking
LLMRanker uses an LLM to reason about relevance rather than rely on similarity scores (like cross-encoders).
This leads to better results for complex queries and helps provide higher-quality context for RAG and agent workflows while keeping context windows lean.

🧠 Discover supported models programmatically
Chat generators now expose SUPPORTED_MODELS variable, so you can easily see which models are supported.

💙 Big thanks to our contributors to this release!

🚀 @GoogleDeepMind just launched Gemini Embedding 2 today and Haystack supports it from Day 0.

Embed text, images, video, audio, and PDFs into one vector space and build multimodal search or cross-modal retrieval systems.

Learn how in our latest blog post 👇
haystack.deepset.ai/blog/multimoda…

When @AIatMeta (Facebook AI back then) released it in 2017, most people hadn't even considered that you could search across millions of vectors in milliseconds. FAISS made it real, and the wave of vector databases that followed owes a lot to what it proved was possible.

Years ding. When your project doesn't need a full-blown vector database - think local development, lightweight deployments, or small to medium datasets - FAISS is often the smartest choice. Fast, dependency-free, and battle-hardened by nearly a decade of production use.

We've now integrated it into Haystack as a Document Store. Drop in FAISSDocumentStore for in-memory or persistent vector search, and pair it with FAISSEmbeddingRetriever to plug directly into your pipelines and agents. GPU support included.

🐍 pip install faiss-haystack
🔗 Documentation: haystack.deepset.ai/integrations/f…

Duplicates and near-duplicates are common in Retrieval Augmented Generation. Even if you try not to repeat yourself, your documents will often contain paraphrases of the same information, which adds no value for the LLM and only clutters your context. This increases token usage useful results that did not rank as highly. Diversification reranks results to keep the most unique ones while preserving relevance.

We didn’t want to reinvent the wheel, so we integrated directly with pyversity by @thomas_v_dongen, a small yet comprehensive Python library that provides many diversification algorithms out of the box, including MMR and DPP. It is now integrated into Haystack as a ranker component that fits into your existing pipelines and agents!

🐍 pip install pyversity-haystack

🔗 Documentation: haystack.deepset.ai/integrations/p…

🚀 Haystack 2.25 is here!

This release focuses on improving the user experience of the Agent component and making tool selection more scalable and efficient.

🛠️ Dynamic Tool Discovery with SearchableToolset
With SearchableToolset, agents dynamically discover relevant tools ch instead of exposing hundreds of tools to the LLM upfront.
This approach:
– Reduces LLM context usage (and token costs)
– Improves tool selection accuracy
– Works especially well when connecting to MCP servers

Agents load only the tools they actually need, making large tool ecosystems practical without overwhelming the model.

📝 Reusable Jinja2 Prompt Templates for Agents
Agents now natively support templated` user_prompt` and `required_variables`, simplifying how you structure and invoke agents.
With this update, you can define prompts once and pass variables at runtime. This means:
– Less boilerplate
– Cleaner agent invocation
– More reusable and maintainable agent workflows

💙 Big thanks to our contributors to this release!

👇 Full release notes in the next post

Explore the release and let us know what you build with it 💬

🧪 Haystack & @mem0ai integration

We've added @mem0ai support to haystack-experimental, bringing persistent memory capabilities to your pipelines.

Mem0 allows your Haystack applications to maintain memory across sessions - storing user preferences, and extracted facts that persipeline runs. This enables truly stateful AI systems that learn and adapt over time.

🔍 Key capabilities:
- Cross-session memory persistence for agents and chatbots
- User and session-level context management
- Seamless integration with Haystack's agents

This integration is experimental - we're looking for early adopters to test it and share feedback!

📖 Cookbook:
haystack.deepset.ai/cookbook/memor…

Haystack 2.24 is live 🚀

Here are the highlights:

🔌 Simpler Pipelines
New improvements in the Pipeline logic simplify the component connections, eliminating the need for OutputAdapter and joiners in many common setups, such as query rewriting and hybrid retrieval. This results pipelines even for complex use cases.

🏁 Rankers handles duplicate documents
Rankers now deduplicate documents on top of ranking, enabling cleaner hybrid retrieval pipelines with less components.

📃 File Support in Chat Generators
You can now attach files (e.g., PDFs) to ChatMessage using the new FileContent dataclass when working with OpenAI and Azure chat generators (Responses API). This enables smoother document-based prompting in deterministic flows and agentic settings.

💙 Big thanks to our contributors to this releaseH

👇 Full release notes in the comments

📃@PaddlePaddle released a new VLM: PaddleOCR-VL-1.5 (0.9B) and you can already use it in your document-heavy Haystack pipelines.

PaddleOCR-VL-1.5 goes beyond classic OCR: it understands document layout and structure, extracting tables, formulas, charts, and key elements from nd images. This makes it a powerful building block for reliable RAG pipelines and document-centric AI applications.

For teams building AI systems for reasoning over complex documents, this enables more accurate retrieval, grounding, and reasoning across document structure.

🔍 Why it’s exciting:
- 94.5% accuracy on OmniDocBench v1.5
- Irregular-shaped localization for real-world documents (skew, warp, photos)
- Strong improvements in table, formula, and text spotting
- Multilingual support, including rare scripts and complex layouts

🔗 Model: huggingface.co/PaddlePaddle/P…
🔗 Docs: haystack.deepset.ai/integrations/p…

🚀 Haystack 2.23 is out!

This release focuses on making agents safer, more visual, and components easier to customize with fewer custom hooks and less boilerplate.

Here are the highlights:

🔄 Human-in-the-Loop for Agents
Agents can now pause for human confirmation before n define confirmation behavior per tool (always ask, ask once, or never ask), ideal for workflows with sensitive operations.

🖼️ Image Support for Tool Results
Tools can now return images alongside text. Agents can retrieve and reason over images, unlocking multimodal workflows like visual search, image inspection, and richer tool outputs.

🧩 Simpler Serialization for Custom Components
Custom components now serialize and deserialize automatically in most cases, even when they include complex objects as parameters. This means less boilerplate, easier pipeline snapshots for recovery, and simpler custom component API.

💙 Big thanks to our contributors for the release

👉 Full release notes: haystack.deepset.ai/release-notes/…

🚀 Haystack v2.22 is out!

This release brings smarter doc chunking, improved component behavior, and easier tool-to-LLM integration with fewer manual steps.

Highlights
✂️ New `EmbeddingBasedDocumentSplitter` for smarter document chunking
🛠️ `outputs_to_string` lets tools with ease
🔥 Components now auto-trigger `warm_up()` on first run
🐍 Python 3.10+ is now required

Big thanks to our contributors 💙

Full release notes👇
🔗 haystack.deepset.ai/release-notes/…

One Name. One Product Family. One Look 💙

We’re unifying the Haystack ecosystem at @deepset_ai under one name and a new logo, reflecting its role as a framework, a community, and the foundation of our enterprise platform.

👉 Read the announcement: haystack.deepset.ai/blog/announcin…

🚀 Haystack 2.21 is out!

🔍 Smarter, Broader Retrieval with Multi-Query RAG

This release introduces three new components that significantly boost retrieval recall in RAG systems by expanding the user query and retrieving documents across multiple reformulations:

- semantically similar variations of a user query to broaden search coverage.

- MultiQueryTextRetriever runs multiple queries in parallel using a text-based retriever (e.g., BM25) and merges results by score.

- MultiQueryEmbeddingRetriever performs the same multi-query retrieval flow using embeddings, enabling richer semantic recall.

Used together, these components create a multi-query retrieval pipeline that improves recall especially when queries are short, or ambiguous.

✨ Plus: more upgrades, enhancements and minor fixes to keep everything running smoothly.

👉 See the full release notes: haystack.deepset.ai/release-notes/…

#haystack #opensource

✈️ Lufthansa Industry Solutions x Haystack

We’re excited to share how @LHIND_DLH built SmartAssistantAI, an enterprise-grade, compliant AI assistant powered by Haystack.

This new story dives into how the LHIND team:

🔹 Uses RAG to make company knowledge easier to access
🔹 g and hybrid search
🔹 Ensures compliance and observability
🔹 Measures success

Plus, some valuable lessons learned along the way.

Read the full story to see how enterprise teams are building reliable AI systems with Haystack 👇
haystack.deepset.ai/blog/lufthansa…

🚀 Haystack 2.19 is out!

Here are the highlights:

🛡️ Resilient Generation with FallbackChatGenerator
Automatically switch between multiple LLMs or settings when one fails due to errors like timeout, rate limit, and server errors, keeping your systems responsive no matter what nd Toolset in Agents
Combine both in a single agent for more flexible tool organization and simpler agent design.

♨️ Faster Agents with Tool Warmup
Pre-initialize tools (like DB connections or models) before first use to reduce startup delays.

✨ Plus: smaller improvements and fixes to keep things running smoothly.

👉 See the full release notes: haystack.deepset.ai/release-notes/…

🇪🇺 Honored to announce with the @EU_Commission that we’re developing the AI@EC Platform and building Europe’s foundation for sovereign, orchestrated AI.

Powered by @Haystack_AI and co-developed for transparency, trust, and EU sovereignty, this marks a major milestone for open en, community-driven technology can power mission-critical AI at the scale of Europe.

🚀 Haystack 2.18 is here!

🔁 Pipeline Error Recovery → resume from the last successful step with snapshots
🧠 Structured Outputs → generate Pydantic/JSON schema responses with @OpenAI/@Azure
🛠️PipelineTool → easily expose full @Haystack_AI pipelines as LLM-compatible tools
t → update Agents dynamically

👉 Check out the full release notes
haystack.deepset.ai/release-notes/…

🔥 Haystack is trending on GitHub… again!

It’s all thanks to you, the devs who use, contribute, and share ideas. You’re shaping how AI apps get built for production.

⭐ Add your star and join the community: github.com/deepset-ai/hay…

🧪 Haystack Experimental v0.13.0 is here!

✨ New experiments:
- Smart document splitting that understands context
- AI that recognizes when it might be hallucinating
- Query expansion that searches multiple angles of the query at once

Try the new experiments: github.com/deepset-ai/hay…

🔔 Haystack 2.17 is here with extended multimodality & flexible routing

This release expands Haystack's multimodal capabilities and makes core components more powerful and flexible for building dynamic AI pipelines.

🖼️ Image support for more model providers
Building on the imagHaystack 2.16, we now support even more providers:
@amazon Bedrock, @AnthropicAI, @Azure, @Google, @huggingface API, @Meta Llama API, @MistralAI, @nvidia, @ollama, @OpenAI, @openrouter, and STACKIT.

🧩 Extended components
• MetadataRouter can now route not only Documents but also ByteStream objects.
• SentenceWindowRetriever is more flexible now: its source_id_meta_field now accepts multiple fields, enabling retrieval only for documents matching all specified meta fields.

📘 Full release notes 👉 haystack.deepset.ai/release-notes/…

🙌 Huge thanks to our contributors and community for driving Haystack forward!

🕵️🌐 New notebook: Browser Agents with Gemini + Playwright MCP

No API? No problem. This Agent can use websites like you do: click, type, wait, read.

🎥 In the video, it visits @huggingface Spaces, finds FLUX.1 [schnell], expands "my holiday on Lake Como" into a detailed prompt, returns the URL.

This agent can also:
🗞️ Compare news sites → table of shared stories with links
▶️ Find content creator social profiles from YouTube videos
🛍️ Check a product's Amazon price range
🚂🚌 Gather public transport travel options

📓 Step-by-step notebook by @theanakin87: haystack.deepset.ai/cookbook/brows…

Built with Haystack + @googledeepmind Gemini 2.5 Flash + Playwright MCP server

#mcp #agent

🔔 Haystack 2.16 is here with deep debugging for Agents and powerful multimodal workflows.

This release is all about developer control and building smarter, more flexible AI apps.

🧠 Agent Breakpoints
Pause execution mid-run, inspect internal state, and resume. Perfect for flows and improving traceability.

🖼️ Multimodal Agents & Pipelines
Haystack now understands text and images.
Build applications that:
• Extract context from scanned PDFs
• Embed image files
• Prompt LLMs with visuals

👉Learn more in the release notes: haystack.deepset.ai/release-notes/…