Charles Pierse

ML @weaviate_io | Occasional maker of things, regular breaker of things.

[Charles Pierse @cdpierse]

Very proud to see Transformers Interpret hit 1k stars today with over 200,000 downloads. Maintaining a library has been one of the hardest and most rewarding things I've done. Looking forward to adding some very cool new explainers **cough** Seq2Seq LLM's models very soon 👀

[Bob van Luijt @bobvanluijt]

🚢 Toot! The team is shipping again!

We’re seeing a massive wave of developers & agents building advanced AI memory capabilities on top of Weaviate. To make that even easier, we just launched Engram, a dedicated memory service built right on our database.

[Weaviate AI Database @weaviate_io]

Most agent memory systems are just glorified context windows.

And this is exactly why production agents fail at scale.

We've been working on this for months, and it's finally here: 𝗘𝗻𝗴𝗿𝗮𝗺 𝗶𝘀 𝗻𝗼𝘄 𝗚𝗔.

If you've been building agentic applications, you know the problem. Agents that should get smarter over time stay flat instead. They forget user preferences, re-solve the same problems repeatedly, and waste tokens on work that can't be reused. Long context windows help, but cramming them full degrades accuracy, inflates costs, and increases latency.

𝗘𝗻𝗴𝗿𝗮𝗺 𝘀𝗼𝗹𝘃𝗲𝘀 𝘁𝗵𝗶𝘀.

It's a managed memory service built on Weaviate that 𝘢𝘤𝘵𝘪𝘷𝘦𝘭𝘺 𝘮𝘢𝘪𝘯𝘵𝘢𝘪𝘯𝘴 memory instead of just storing it. Asynchronous pipelines extract relevant information from raw data, reconcile it with existing memories (handling deduplication, preference changes, time-evolving facts), and persist clean, structured memory state ready for retrieval.

𝗞𝗲𝘆 𝗰𝗮𝗽𝗮𝗯𝗶𝗹𝗶𝘁𝗶𝗲𝘀:

𝗙𝗶𝗿𝗲-𝗮𝗻𝗱-𝗳𝗼𝗿𝗴𝗲𝘁 𝗔𝗣𝗜 → Add raw data and continue working. Pipelines run asynchronously in the background with durable execution.

𝗧𝗼𝗽𝗶𝗰𝘀 𝗮𝘀 𝗺𝗲𝗺𝗼𝗿𝘆 𝗺𝗮𝗴𝗻𝗲𝘁𝘀 → Natural language descriptions that pull matching information from raw data. You control what's worth remembering.

𝗦𝗰𝗼𝗽𝗲𝘀 𝗳𝗼𝗿 𝗶𝘀𝗼𝗹𝗮𝘁𝗶𝗼𝗻 → Project-wide, user-scoped, or property-scoped memories with hard and soft isolation enforced at the platform level.

𝗖𝗼𝗺𝗽𝗼𝘀𝗮𝗯𝗹𝗲 𝗽𝗶𝗽𝗲𝗹𝗶𝗻𝗲𝘀 → Extract, transform, buffer, and commit steps that manage memories dynamically based on data type and preferences.

𝗕𝘂𝗶𝗹𝘁 𝗼𝗻 𝗪𝗲𝗮𝘃𝗶𝗮𝘁𝗲 → Memory retrieval inherits Weaviate's vector + keyword + metadata search on the same production stack you already trust, using native multi-tenancy to isolate instances.

Whether you're building chatbots that remember user preferences, agents that learn from experience, or multi-agent systems that need shared context, Engram gives you memory as infrastructure.

As a promotional offer, we’re giving $75 in credits for your first three months of Engram! Sign up before July 15th to claim it.

Read the blog: https://t.co/24koipQIjO

Get started: https://t.co/v9fMGMNCej

[Connor Shorten @CShorten30]

Introducing Suggest Queries Mode 💚

Weaviate’s Query Agent can now suggest queries that are answerable by the data in your collections.

This helps users discover what kinds of questions they can ask! 💬

[Charles Pierse @cdpierse]

@victorialslocum @etiennedi @bcherny 🤣🤣🤣🤣🤣

[Victoria Slocum @victorialslocum]

𝗨𝗻𝗽𝗼𝗽𝘂𝗹𝗮𝗿 𝗼𝗽𝗶𝗻𝗶𝗼𝗻: The best hires aren't experienced.

They're hungry. And we're searching for two of them right now.

We have two open positions in the Growth team at @weaviate_io:

𝗗𝗲𝘃𝗲𝗹𝗼𝗽𝗲𝗿 𝗚𝗿𝗼𝘄𝘁𝗵 𝗜𝗻𝘁𝗲𝗿𝗻
𝗗𝗲𝘃𝗲𝗹𝗼𝗽𝗲𝗿 𝗠𝗮𝗿𝗸𝗲𝘁𝗶𝗻𝗴 𝗠𝗮𝗻𝗮𝗴𝗲𝗿

We're looking for people who aren't afraid to work hard, who want to grind, and who are hungry to learn how to define narratives and win on social media.

Not people who check boxes. Not people who want a comfortable 9-5. We want the next diamonds - people who see an opportunity to learn from a team that's built a developer community from the ground up and think "I want to know how they did that."

You'll learn:
• How to craft narratives that cut through noise
• What actually works on social media (not what the courses tell you)
• How to build genuine engagement with developer communities
• The difference between marketing 𝘢𝘵 developers and marketing 𝘸𝘪𝘵ℎ them

This isn't going to be easy. But if you're the type of person who gets excited by that statement rather than intimidated by it, you might be exactly who we're looking for.

Apply here: https://t.co/v0lU0QfX7B
Complete the challenge. Impress us. 💎

[Connor Shorten @CShorten30]

How do we build search systems for Agents? 👾🔎

I am SUPER EXCITED to share a new episode of the Weaviate Podcast with Zijian Chen (@zijian42chen) and Xueguang Ma (@xueguang_ma) from the University of Waterloo on AgentIR! 🎙️💚

When humans search, we write short queries and keep our reasoning in our heads. Deep Research agents do the opposite. They leave reasoning traces that reflect on prior results, clarify intent, and plan what to search next. Existing retrievers completely ignore this signal because they were designed for human queries. 💭

AgentIR jointly embeds the agent's reasoning trace alongside its query, training a retriever that actually understands what the agent is thinking. AgentIR-4B hits 68% accuracy on BrowseComp-Plus compared to 52% for conventional embedding models twice its size. 📊

One idea I found especially interesting is how AgentIR raises context management questions for agents: what should be remembered, compacted, or retrieved just in time? The current reasoning trace naturally curates history by summarizing confirmed findings and filtering out wrong guesses. Forgetting becomes a feature, not a bug. 🔬

We also covered BrowseComp-Plus, their benchmark for disentangled evaluation of agents and retrievers, and the open question of scaling search deeper vs. wider.

If you're working at the intersection of Agents and Search, I think you'll get a lot out of this one! Links below! 🎉

[Charles Pierse @cdpierse]

We've been tinkering away on something new at @weaviate_io about the role memory plays not just for chatbots but for the agents you have deployed in prod.

It's become clear that memory isn't a nice to have but that it's essential layer for you apps and agents to succeed in the long run.

[Weaviate AI Database @weaviate_io]

Your AI agent worked perfectly in January.

By June, it's confidently giving you wrong answers. Here's why:

As AI applications graduate from PoCs to production, we're hitting a wall that better models can't solve: 𝗹𝗮𝗰𝗸 𝗼𝗳 𝗰𝗼𝗻𝘁𝗶𝗻𝘂𝗶𝘁𝘆.

𝗧𝗵𝗲 𝗹𝗶𝗺𝗶𝘁𝗲𝗱 𝗹𝗼𝗼𝗽 𝗽𝗿𝗼𝗯𝗹𝗲𝗺
Today's AI applications treat each interaction as largely disposable. You've felt it already: repeating preferences, restating context, and re-teaching the same facts.

At agent scale, the problem worsens. Agents re-derive the same conclusions, regenerate identical facts, and discard half-finished work, and what looks like forgetfulness for humans turns into systemic chaos for machines.

𝗪𝗵𝘆 𝗻𝗮𝗶𝘃𝗲 𝗺𝗲𝗺𝗼𝗿𝘆 𝘄𝗶𝗹𝗹 𝗳𝗮𝗶𝗹
Here's what happens with a basic memory implementation:
Week 1: Magic! The agent remembers.
Month 3: Responses slow down as memory bloats.
Month 6: Answers drift wildly as the model pulls from conflicting and outdated context.
Helpful continuity has slowly turned into accumulated noise.

𝗧𝗵𝗲 𝘀𝗵𝗶𝗳𝘁: 𝗺𝗲𝗺𝗼𝗿𝘆 𝗶𝘀𝗻’𝘁 𝘀𝘁𝗼𝗿𝗲𝗱, 𝗶𝘁’𝘀 𝘮𝘢𝘪𝘯𝘵𝘢𝘪𝘯𝘦𝘥.
Useful memory systems actively manage context through write control, deduplication, reconciliation, amendment, and purposeful forgetting.

Without these, memory becomes an ever-growing pile of notes. With them, it becomes 𝗿𝗲𝗹𝗶𝗮𝗯𝗹𝗲 𝘀𝘁𝗮𝘁𝗲.

At Weaviate, we treat memory as a first-class data problem: durable, governable, and safe under change.

Read the full blog post on our vision for memory and signup for the product preview: https://t.co/fmC9r6wMgP

[Charles Pierse @cdpierse]

When we were building Weaviate's Transformation Agent out, we knew we needed to saturate our GPUs as much as possible and squeeze every bit of utilization we could. Not only was this possible with @modal but it made the whole process of implementation incredibly straightforward.

[Modal @modal]

It's time to run your own LLM inference. The open models and open source engines are ready. Are you?

We've been working with leading teams like @DecagonAI, @weaviate_io, and @reductoai to ship production-grade inference.

Here's how we do it:

[Connor Shorten @CShorten30]

Search Mode is now available in the Weaviate Console! 💻🎨

A new button lets you toggle between Ask and Search! 🖱️

Another cool aspect of this is that you can switch to Search Mode in the middle of a conversation with Ask Mode! 💬

Check it out! 👇

[Philip Vollet @philipvollet]

Multi-vector models like ColBERT and ColPali are incredible for retrieval quality, but they eat memory for breakfast. We're talking 10-13x more memory than single-vector embeddings because you're storing hundreds of vectors per document instead of one.

MUVERA (Multi-Vector Retrieval via Fixed Dimensional Encodings) changes this by encoding multi-vector embeddings into single fixed-size vectors through space partitioning and dimensionality reduction.

Real results from our LoTTE benchmark tests:

- Memory footprint: 12GB → <1GB (~70% reduction)
- Import speed: 20+ minutes → 3-6 minutes
- Recall: 80-90%+ with proper HNSW ef tuning

The trade-off? Some recall loss that you can mitigate with higher ef values, though that reduces query throughput. But when you're looking at tens or hundreds of thousands of dollars in cloud costs per year, this is a no-brainer for large-scale deployments.

Available in Weaviate 1.31+ and it's literally just a couple lines of code to enable!

[Leonie @helloiamleonie]

I've been learning about agent memory for the past few weeks.

This new blog summarizes everything I've learned so far:

• What is agent memory, and why do you need it
• What are the types of memory (and what categorization approaches are there?)
• How do you manage memory in AI agents

Join me in learning about memory for agents:
https://t.co/iVTRKAp3Fz

[Connor Shorten @CShorten30]

I am SUPER EXCITED to publish the 130th episode of the Weaviate Podcast featuring Xiaoqiang Lin (@xiaoqiang_98), the lead author of REFRAG from Meta Superintelligence Labs! 🎙️🎉

Traditional RAG systems use vectors to retrieve relevant context, but then throw away the vectors, just giving the content to the LLM. REFRAG instead feeds the LLM these pre-computed vectors, achieving massive gains in long context processing and LLM inference speed! 🧬

REFRAG makes Time-To-First-Token (TTFT) 31x faster and Time-To-Iterative-Token (TTIT) 3x faster, boosting overall LLM throughput by 7x while also being able to handle much longer contexts! 🔥🔥

There are so many interesting aspects to this and I loved diving into the details with Xiaoqiang! I hope you enjoy the podcast! 🎙️

[Charles Pierse @cdpierse]

@jackhodkinson93 And so goes the cycle ➰

[Connor Shorten @CShorten30]

As a quick TLDR, there are two key aspects to understanding how REFRAG works:

1. The particular way REFRAG represents context tokens and injects them into LLM decoding, as well as how this speeds up LLM inference. ⚡️

2. The training algorithm used to align the encoder, projection layer, decoder, and selective chunk expansion policy! 🏭

[Weaviate AI Database @weaviate_io]

We benchmarked the Query Agent’s Search Mode vs. Hybrid Search across 12 IR benchmarks from BEIR, LoTTe, BRIGHT, EnronQA, and WixQA.

The results? +17% average improvement in Success @ 1 and +11% in Recall @ 5!

Learn more about the benchmarks and dive into our experimental details:

📊 Blog post: https://t.co/8FGLNIhTxX

[Clint J. @SearchDataEng]

I just finished reading @weaviate_io new blog on Search Mode benchmarks , and it’s a real milestone for the vector search + LLM community.

Search Mode is Weaviate’s new compound retrieval system , basically a smarter way of doing search that goes beyond “hybrid” (keyword + vector).

To go beyond standard hybrid search , Weaviate has developed several important methods .

Query Expansion :
Adds related terms so the engine doesn’t miss relevant docs.

Query Decomposition :
Breaks a complex user request into smaller parts.

Reranking :
Intelligently reorders results so the best answers rise to the top.

Think of it as search that not only finds but also understands how to shape the question, and the response!

Benchmarks tested :

1. BEIR (standard IR benchmark across multiple datasets)
2. LoTTe (long-tail, hard questions)
3. BRIGHT (biomedical retrieval)
4. EnronQA (enterprise email data)
5. WixQA (real-world domain-specific QA)

On every one of these, Search Mode shows gains over hybrid search!

As Weaviate is increasingly turned to as the preferred Search engine for mission critical applications, there are important benefits.

For RAG pipelines, higher precision means fewer hallucinations.

For agentic systems, decomposition means the retriever can keep up with multi-step reasoning.

For enterprises, messy datasets (emails, docs, portals) become far more searchable without manual tuning.

Amazing, absolutely amazing!

[Connor Shorten @CShorten30]

I am SUPER excited to share our new Information Retrieval benchmarks! 🥳

Search Mode is a Compound Retrieval System that utilizes Query Expansion, Query Decomposition, Reranking, and more to achieve super accurate search results! 🎯

The blog post demonstrates how it performs compared to Hybrid Search on the BEIR, LoTTe, BRIGHT, EnronQA, and WixQA benchmarks! 📊

We also describe further what these benchmarks are and why we chose them! I hope you find this interesting! 👇👇

[Connor Shorten @CShorten30]

I am SUPER excited to publish the 128th episode of the Weaviate Podcast featuring Charles Pierse (@cdpierse)! 🎙️🎉

Charles has lead the development behind the GA release of Weaviate's Query Agent! 🔎

The podcast explores the 6 month journey from alpha release to GA! Starting with the meta from unexpected user feedback, collaboration across teams within Weaviate, and the design of the Python and TypeScript clients. 👋🤝🎨

We then dove deep into the tech! Discussing citations in AI systems, schema introspection, multi-collection routing, and the Compound Retrieval System behind search mode. 🛠️

Back into the meta around the Query Agent, we ended with its integration with Weaviate's GUI Cloud Console, our case study with MetaBuddy, and some predictions for the future of the Weaviate Query Agent! 🚀

I had so much fun chatting about these things with Charles! I really hope you enjoy the podcast! 🎙️

[Charles Pierse @cdpierse]

@CShorten30 It was so great to be back on again Connor !

[Charles Pierse @cdpierse]

We've been heads down the past few month getting the Query Agent ready for graduation to GA 🎓

Super proud of all the work from the team at @weaviate_io that has gone into getting it to where it is today.

If you want to supercharge your retrieval or build a complex chatbot on top of your Weaviate data, its a simple as a single line of code with Query Agent!

[Weaviate AI Database @weaviate_io]

We’re excited to announce:
The Weaviate Query Agent is now GA!

WQA is a Weaviate-native agent that transforms natural language questions into precise database operations, giving you reliable, fully transparent results.

It supports:
• Dynamic filters
• Smart routing across collections
• Aggregations
• Accurate results with full source citations

The result? Faster, more reliable, and fully transparent data-aware AI.

How to get started:
• 𝗜𝗻 𝘁𝗵𝗲 𝗖𝗼𝗻𝘀𝗼𝗹𝗲: Explore your data with natural language. See the Agent's "thought process" and the sources it used.
• 𝗩𝗶𝗮 𝗔𝗣𝗜𝘀/𝗦𝗗𝗞𝘀: Embed this intelligent querying directly into your applications, reducing boilerplate and shipping faster.

Key benefits:
• Say goodbye to custom query-rewriting pipelines.
• Get structured, predictable data back.
• Full transparency: see every filter, aggregation, and source.
• Automatically handles queries across multiple data collections and tenants.

Try it yourself:
🔬 Colab quickstart → https://t.co/Z7ZMDUVwTf
✍️ Launch blog → https://t.co/wJHrZ70Lca