AI infra for devs. RAG, vector search, AI agents, and the database behind them. Follow for practical experiments, benchmarks, architecture and code.
Joined January 2026
- Tweets1.2K
- Following24
- Followers175
- Likes767
Pinned Tweet
We grew a Pi 5 agent's memory 100x, to 100,000 vectors. Median recall moved from 1.90 ms to 4.17 ms. A 2.2x cost for 100x the data, all on the board. Here's how to do it.
How VectorAI DB handles it:
• Scalar quantization: float32 to int8, ~4x smaller
• Product quantization: smaller still
• IVF indexing when HNSW won't fit in memory
Pick the memory, recall and latency trade-off per deployment 👇
actian.com/blog/databases/ac…
RAM prices keep climbing, and your vector index lives in RAM.
1M vectors × 1024 dims:
float32: ~4.1 GB
int8: ~1 GB
4x less memory. Rescore the top hits at full precision and you keep most of the recall.
Even @perplexitydevs ships int8 embeddings now 👇
teapawt.net/perplexitydevs/status/…
RAM prices keep climbing, and your vector index lives in RAM.
1M vectors × 1024 dims:
float32: ~4.1 GB
int8: ~1 GB
4x less memory. Rescore the top hits at full precision and you keep most of the recall.
Even @perplexitydevs ships int8 embeddings now 👇
teapawt.net/perplexitydevs/status/…
Introducing pplx-embed-v2-context: open-weight vision-language embedding models in 9B and 0.6B.
Both return one int8 vector per document, or one per chunk that sees the whole document.
Read the paper, written with @turbopuffer: arxiv.org/abs/2610.38854
Help us map the self-hosted AI stack 🗺️
We've listed 179 projects you can run on your own hardware. What did we miss?
Make a PR!
bki.sh/uLHircK
Join us in the self-hosted-ai channel on Discord
bki.sh/mVvsrBo
Actian for Developers retweeted
If I had 6 months to become an AI engineer
M1 → Python, APIs, Git
M2 → LLM basics: tokens, context, prompting
M3 → RAG: embeddings, vector DB, chunking
M4 → evals and tracing
M5 → agents with tool calling
M6 → ship 2 projects with live demos
Proof beats certificates
Chroma Cloud users, your house hunt has a deadline. Give your vectors a home bki.sh/W3cGtG8
Chroma Cloud is moving out. Time to find your vectors a new home 🫂 bki.sh/KE03HfT
We grew a Pi 5 agent's memory 100x, to 100,000 vectors. Median recall moved from 1.90 ms to 4.17 ms. A 2.2x cost for 100x the data, all on the board 🫡
Self-hosted mem0 still sends your agents’ conversations to OpenAI by default.
Out of the box, it calls OpenAI to extract facts and again to embed them, even when you host it yourself.
Swap in Ollama and VectorAI DB and your agents’ memory runs entirely on your own hardware.
We confirmed it with the network blocked.
Here’s how to run it fully offline, too 👇
bki.sh/RGU6Glg
Actian for Developers retweeted
What if your AI coding agent fixed a bug today, but then forgot the solution in the next session?
Came across a GPT-6 Astra experiment where they used Actian VectorAI DB as external memory to store and retrieve previously tested fixes across sessions.
It worked by:
-Search a vector DB for similar previous fixes
-Check if the old fix applies to the current code
-Save new fixes only after tests pass
They also tested it across 5 sessions and the agent successfully retrieved earlier fixes. I liked the idea of remembering tested fixes instead of the whole conversation.
I'm curious what others think about giving AI coding agents this kind of memory. Would it be useful in your workflow?
Half the AI timeline right now: "You don't need a vector database."
The other half: "you'll regret not having one."
Both are right, depending on your workload.
So we built a tool that tells you which camp you're in.
Answer a few questions about your data, latency, filtering and deployment.
In 30 seconds, you'll know whether a flat file, keyword search or pgvector is enough, or whether you really need a vector database.
Sometimes the answer is no. We'll tell you that too.
bki.sh/RyqhiNN
100% state-owned. Air-gapped. On-premises. 🔐🇮🇳
India is drawing its sovereign AI perimeter.🛡🧱
The Centre is formalising a managed-service model for critical-sector AI.
Defence, healthcare, banking and other strategic workloads will no longer run on rented commercial cloud.
The government will own the physical GPU clusters and data-centre infrastructure on its own premises.
Private operators may only manage the workloads, under strict non-disclosure and security frameworks.
Why this is a structural shift:
• CERT-In wake-up call: Recent evaluations found rogue AI agents overstepping their mandates and querying balance sheets and sensitive records across Indian commercial banks. Virtual tenancy on third-party clouds is no longer considered safe for strategic assets.
• From subsidy to ownership: CDAC’s twin RFPs: one for state-owned GPU capacity, one for dedicated high-density facilities, signal New Delhi is moving beyond compute subsidies toward direct asset control.
• Three-tier security: Mission-critical workloads stay fully air-gapped and government-owned. Less sensitive categories may use a managed-service layer where the state still owns the silicon. Public commercial compute continues under MeitY’s MeghRaj framework.
• Algorithmic containment: If domestic foundation models train on imported hardware or sit overseas, policy and defence levers remain exposed to foreign jurisdictions.
Sovereign AI cannot run on rented digital real estate.
Controlling the physical silicon layer is the only durable defence against technological leverage.
@AshwiniVaishnaw @cdacindia @IndianCERT
By @SurabhiA_ET @Subhayan_ET
#SovereignAI #DeepTech #CyberSecurity #IndiaAI #Semiconductors #NationalSecurity #ViksitBharat
Use this hack to fix a vector search that's silently returning the wrong results bki.sh/ujXWsUV
A filter that returns a sixth of the right results and reports success is worse than one that crashes. That's what a metadata filter does to recall on a graph index at default settings, and almost nobody catches it. bki.sh/ujXWsUV View article
A filter that returns a sixth of the right results and reports success is worse than one that crashes. That's what a metadata filter does to recall on a graph index at default settings, and almost nobody catches it. bki.sh/ujXWsUV View article
What if we built a memory layer for your @crewAIInc agents? We did!
The default memory backend can show one user’s private memories to another, throw “database is locked” errors under concurrent load, and lose its store when a container restarts without a volume mount.
Our replacement runs on VectorAI DB. It’s one new file and two changed lines in your Crew setup, and your agents and tasks stay as they are.
We tested all three failures and put the walkthrough on video.
