Elastic Adds jina-ocr-v1 Document Parser and Technical Preview of Elastic CLI

jina-ocr-v1: layout-aware parsing in one model

According to Elastic, jina-ocr-v1 takes scanned pages, document photos, presentation slides and handwritten notes and returns structured Markdown in a single pass, without a pipeline of separate specialist models. Tables are output as HTML and mathematical formulas as LaTeX. Text from embedded images such as book covers and logos is recognized but left out of the output.

Elastic says the model understands page layout first and then reads in the order a human would, whereas standard OCR reads characters left to right and often scrambles multi-column layouts and sidebars. The model has 3.4 billion parameters in a mixture-of-experts design, with only 570 million active at inference time.

On olmOCR-bench, jina-ocr-v1 scores 83.4, which Elastic describes as the highest of any model with fewer than 600 million active parameters. On OmniDocBench it scores 91.14, above GPT-5.2 at 86.59, according to the company. Models that beat it on both benchmarks have significantly more active parameters.

Language support covers more than 100 languages, including Arabic, Chinese, Japanese, Korean, Thai, Hindi, Greek, Cyrillic, Turkish and Czech. Mixed-language pages are handled in one pass without extra configuration.

• Elastic Inference Service: available now via the chat_completions inference endpoint, with native PDF support coming soon.
• Jina API: standard HTTP access with 10 million free tokens to get started.
• Local or on-prem: download from Hugging Face under CC BY-NC 4.0, or use Jina On-Prem for commercial deployments.
• Jina AI Reader API: add X-Respond-With: jina-ocr-v1 to the request header; PDFs and HTML pages are converted to images automatically.

Elastic CLI and rebuilt Agent Skills

The Elastic CLI is in technical preview and offers a single entry point for public Elasticsearch, Kibana and Elastic Cloud APIs. It installs with npm install -g @elastic/cli. Every command takes JSON in, via stdin or --input-file, and returns JSON out with --json. Inputs are validated against a JSON Schema before any request is sent, and failures return a non-zero exit code.

Several features target use with AI agents:

• API keys are stored in the OS keychain (macOS Keychain, Linux Secret Service or Windows Credential Manager), so they do not appear in shell history or LLM transcripts.
• Allowlists and blocklists restrict which commands can run, globally or per context, for example to give an agent read access but not delete access.
• Destructive commands require confirmation in a terminal. In non-interactive sessions they only run if --yes is passed.
• --output-fields and --output-template trim responses to keep agent context windows small.

The CLI also includes helpers for bulk ingest, scroll search and watching new documents.

Elastic Agent Skills, which teach coding agents such as Claude Code, Cursor, Copilot and Gemini CLI to work with Elasticsearch and Kibana, now use the CLI as their transport. Previously they hardcoded details like curl commands with embedded API keys and hostnames, which break when credentials rotate or endpoints change. Skills for Elastic Cloud onboarding, Elastic Workflows and Kubernetes investigation are available now. Skills for ES|QL, ingest, reindex, index design, Kibana dashboards and alerting are in progress.

Knowledge Indicators and LLM tracing

A post from Kathleen DeRusso and team describes an alternative to standard retrieval augmented generation (RAG), which loads full documents into an agent's context for each question. Instead, a Kibana Workflow runs a one-time batch pass over a corpus. An LLM extracts a Knowledge Indicator for each document: a compact summary with a title, plain-language summary, entities, questions the document can answer, and a tagline. These are written to an Elasticsearch AI Index, and agents run a hybrid BM25 and semantic search over them with ES|QL.

In one Wikipedia test query, about the actress who played Torvi in Vikings, Elastic reports standard RAG used 8 tool calls, 386,187 tokens and 44.86 s of latency. With Knowledge Indicators it used 2 tool calls, 27,625 tokens and 15.22 s. The newsletter says the answers were equivalent and grounded in the same source material, and puts the overall gain at 93% fewer tokens and a third of the latency. The approach works with LangChain, Elastic Agent Builder and Claude Code without changes to the agent harness.

Separately, LLM spans in Elastic APM now show prompts, responses and per-call token counts in the trace waterfall for applications instrumented with OpenTelemetry. Elastic Security Labs also instrumented 1,100 developer machines with Cursor hooks and collected 13 million tool call events.

What it means for developers

For teams handling mixed document collections, jina-ocr-v1 offers a single model route to structured output, with options from a managed service to on-prem deployment. Note that the Hugging Face download is licensed CC BY-NC 4.0, so commercial use goes through Jina On-Prem.

The CLI and updated Agent Skills aim to let agents act on Elastic deployments with credential protection and policy limits. The Knowledge Indicator figures come from a single test query, so results on other workloads may differ. Details, code and benchmark tables are in Elastic's linked posts.