Haystack

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 added

And this is just Day 1. New drops every day this week at 3 PM CEST.
Full announcement 👇
haystack.deepset.ai/blog/haystack-…

Haystack 3.5 is out 🚀

🧭 New OpenAI Decisions API components route text and images and classify documents.

🧱 StructuredOutputHook for agents validates the final reply against a Pydantic model or JSON schema and retries if it fails.

⏳ Step budget reminders. Agents get told how many steps are left, so they finish with an answer instead of being cut off.

🎯 Multivector embeddings. A new multivector_embedding field and MaxSimRanker support ColBERT-style models.

⚠️ Heads-up: gpt-5-mini shuts down Dec 11. The new default is gpt-5.6-terra, so update any YAML pipelines that still use gpt-5-mini.

Thanks to our 69 contributors for this release 💙

💻 pip install -U haystack-ai
Release notes 👉 haystack.deepset.ai/release-notes/…

Join us in Toronto!

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 older tool results to a 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/…