Elastic

Where developers learn, build, and share. Your source for hands-on demos, cheat sheets, explainers and more.

[Elastic @elastic]

Your AI agent needs answers from millions of events that won't fit in its context.

“Which regions had >5% checkout failures yesterday, with ≥1,000 attempts?”

Give it Elasticsearch aggregations to group, count, calculate, and filter.

The agent gets a compact result.

[Elastic @elastic]

Reference docs: https://t.co/YpSBewpwmS

[Elastic @elastic]

Your API's average latency is 80ms. Looks fine.

Now look closer: 98 requests took 20ms. 2 took 3 seconds.

Averages hide the tail. And the tail is real users, often your heaviest ones, because more requests means more chances to hit it.

Watch p95 and p99, not just the mean.

In Elastic APM, switch the latency chart from Average to 95th or 99th percentile and see who's actually waiting.

[Elastic @elastic]

Reference docs: https://t.co/cTRhSozT4H

[Elastic @elastic]

Ask semantic search for "error E1042" and you may get docs about errors in general.

Not the one with E1042.

Vectors capture meaning. Exact tokens like error codes, SKUs and names get blurred.

Keyword search (BM25) nails exact matches but misses meaning.

So run both, and merge the two ranked lists with Reciprocal Rank Fusion (RRF). Docs that rank high in either list rise to the top, with no need to normalize two very different score scales.

That's hybrid search.

[Elastic @elastic]

Reference docs: https://t.co/8qeh7vJWrH

[Elastic @elastic]

How does search find one word in a billion documents in milliseconds?

It doesn't read the documents.

At index time, Elasticsearch builds an inverted index: every word mapped to the list of documents that contain it. Like the index at the back of a book.

Search "timeout" > jump to that entry > get the doc IDs > fed into BM25 for scoring.

The work moves from search time to write time. That trade is why search is fast.

[Elastic @elastic]

Meta's Muse runs a 37.8M-param reranker on 2 CPU cores.

@rohanadwankar tore down the Muse agent VM. In its memory stack: jina-reranker-v1-turbo-en, 153 MB on disk.

Muse searches its users' memories in Postgres with vectors plus keywords. The Jina reranker orders those hits before the agent reads them.

Per the teardown, retrieval runs inside the VM, so memory chunks stay on the encrypted volume.

It's our English-only reranker from 2024. It still earns its place on a box with no GPU.

You can also run our latest @JinaAI_ models with Jina On-Prem. For self-contained and air-gapped deploys, with no outbound calls. .

Rohan's full teardown: https://t.co/W18dgq9Nt8

[Elastic @elastic]

Did you know you can do K8s analysis with ES|QL on Elastic Observability?

Bookmark this: https://t.co/pissdFyXpF

[Elastic @elastic]

Your alerts are data now. So query them like data.

Kibana Universal Alerting is GA on Elastic Cloud Serverless.

One ES|QL query shows which 10 rules created the most alerts this week. Then you tune them.

- Alert and recovery delays keep a short spike pending instead of alerting
- Action policies route and throttle across rules: one edit changes who gets paged for 100 rules

It runs alongside Classic Alerting. Nothing migrates automatically, so move one rule at a time.

Full walkthrough with the rule YAML and queries: https://t.co/jyQRbtFGwM

[Elastic @elastic]

When a detection rule disappears, the first question is who deleted it.

The second is when, and in which space.

Elastic Cloud Serverless can now ship its audit trail to a Security project in the same organization. Log delivery is generally available, and 8 alerting rules and 7 dashboards arrive with the data.

The dashboards include:
• Kibana Object Forensics: who created, changed or deleted a saved object, and in which space
• Elasticsearch Data Access: who read which index, and who was denied
• Identity & Access: failed sign-ins by credential type, with API keys tracked separately from people

For security and compliance teams, the evidence lands where you already investigate, encrypted in transit and at rest.

5 ignore filters drop noisy events before delivery, so you don't pay to store UI clicks nobody will read.

Filter guide and a closer look at 3 of the dashboards in the blog: https://t.co/m1Xy0NxkfP

[Elastic @elastic]

Full walkthrough, code and setup steps: https://t.co/Qc7hob5Lb4

[Elastic @elastic]

Ask a 56-page scanned PDF a question in plain language.

Get an answer stamped with the page it came from.

jina-ocr-v1 turns every page into structured Markdown, tables included. Elasticsearch embeds it at ingest, so there's no chunking or embedding code to write.

Walkthrough linked below 👇

[Elastic @elastic]

Searching Claude's plugin directory for Elastic used to return nothing.

Now you'll find a plugin with nearly 30 skills across Elasticsearch, Kibana, Security and Observability.

In the video: install the Elastic CLI, add the plugin, then ask Claude to find an error spike in web logs and build a Kibana dashboard around it.

Install it from Settings > Plugins > Discover and search "elastic".

What it took to get into Claude's directory, including the 3 review warnings and why we stopped pointing at main: https://t.co/SL5DLybo8E

[Elastic @elastic]

Tommaso Teofili and Thomas Veasey break down the cost model and error estimation: https://t.co/LMvCtzffHx

[Elastic @elastic]

Opting into DiskBBQ auto calibration raised average QPS 16.7%.

That's indices with auto_calibrate: true vs DiskBBQ's fixed defaults, across 18 public benchmarks.

Quantization bits, oversampling, preconditioning: the right mix depends on your dataset and embedding model, and it shifts every time segments merge.

With auto calibration on, Elasticsearch samples vectors at merge time and predicts recall @ 10 for each candidate encoding and rerank depth.

It stores the most efficient configuration predicted to hit 90% recall @ 10 in that segment's metadata, then re-evaluates on the next merge.

Opted-in indices vs the fixed defaults:

- QPS up 16.7% on average, and up on 15 of 18 datasets
- Over 50% higher QPS on FiQA GTE, Fashion-MNIST and GloVe-200
- Recall up on 15 of 18, including +296% on Fashion-MNIST
- Where either metric dipped, QPS stayed within 1.5% and recall within 2%

The trade-off is at index time. Overhead was under 2% on more than half the datasets and at most 11% on large ones like DBPedia-Entity and HotpotQA. Small datasets saw more, up to 35%, because the calibration sample is a fixed size.

Jina v3 embeddings quantize well enough that calibration chose 1-bit documents on all 3 corpora we tested.

Opt in with one line in index_options:

"type": "bbq_disk", "auto_calibrate": true

[Elastic @elastic]

You have a specific visual in your head, but have no idea which file it’s in.

This walkthrough uses Jina embeddings + Elasticsearch to search video by describing what you’re looking for, in natural language: https://t.co/F0Nxbr9zn7

You receive back the matching scene with timestamps.

[Elastic @elastic]

Proud to make it official with the @audif1_ !

Our relationship started with the team's engineers using Elastic to search and investigate issues across their connected engineering applications. Now, we're building on that foundation through a multi-year partnership, with plans to bring those capabilities to even more applications.

Here's to the seasons ahead and everything we'll build together.

[Audi Revolut F1 Team @audif1_]

New name on the R26 for Austin. 🇺🇸
Welcome, Elastic. Official Race Partner this weekend, Official Search and Observability Supplier from 2027.

[Elastic @elastic]

Elastic is now an Official Supplier of the @audif1_

Behind every engineering breakthrough is a complex, distributed software environment. To keep technical teams moving fast and focused, organizations require immediate visibility into the software and systems that power their engineering work.

Elastic is helping the Audi Revolut F1® Team understand what is happening across its engineering software stack to find signals that matter and act faster. The goal is one operational view from factory to track.

The team already uses Elastic technology as the foundation for its central software logging platform. Starting with the 2027 season, Elastic will become its Official Search and Observability Supplier.

More on the partnership: https://t.co/t6RT9UGNGj

[Elastic @elastic]

When people say “hybrid search,” they usually mean keyword + semantic.

But what about semantic + geospatial?

If I search for “a quiet café to work from,” understanding what I mean is only half the job. I also need it to be nearby.