LlamaIndex launches Extract v2.5 with higher accuracy and grounding
LlamaIndex announced the release of Extract v2.5 on October 1, 2026. The new version improves the F1 score of its three extraction tiers on the ExtractBench benchmark, moving the Cost Effective tier from 87.1 to 93.9, the Agentic tier from 89.8 to 95.8, and the Agentic Plus tier from 95.1 to 96.4. Grounding scores also rose sharply, reaching 80.6 for Agentic and 82.2 for Agentic Plus.
What you can do with it
AI at Work →Audit compliance records with LlamaIndex Extract v2.5
Extracts fields from PDFs, spreadsheets and scanned forms with citations.
- Who for
- marketers
- Cost
- Price not stated
- Effort
- minutes
Use it for
- Extract contract terms
- Populate databases
- Audit compliance records
How to set it up · From the article
- Upload the document to LlamaCloud
- Select Extract v2.5 and enable citations in the extraction config
- Run the job and review the extracted records
Watch out Requires schema design and prompt adjustments
Key points from the news
- ExtractBench overall F1 rose to 93.9 (Cost Effective), 95.8 (Agentic) and 96.4 (Agentic Plus)
- Grounding scores jumped to 80.6 (Agentic) and 82.2 (Agentic Plus) with bounding‑box citations
- Spreadsheet mode lets agents extract directly from workbook cells without extra cost
The upgrade adds a purpose‑built agent harness that can cross‑reference multiple pages, handle complex layouts, and provide bounding‑box citations for difficult fields. It supports long‑list tasks, cross‑page records, scanned forms, and a new spreadsheet mode that works directly with workbook cells. The tool is available through LlamaCloud, a Go‑based CLI, an MCP server, and the Python llama‑cloud SDK, with no increase in per‑page pricing.
LlamaIndex encourages users to test v2.5 on representative documents using their existing schemas, noting that the higher accuracy should reduce manual cleanup and make more workflows fully automated.
Model pages: Sonnet 5.5 → · GPT-6 Sol → · Opus 5.5 →
LlamaIndex Launches Extract V2.5 With Accuracy and Grounding Gains
Unite.AI · 1 October 2026
Loading the full article…
This text was published by Unite.AI and written by Jonas Reeve, Cognitive AI & AGI, AI Research Agent. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Agents & Tools
All →- Decagon launches Voice 3, Chord model, and personal agent gateway · 1 src
- NVIDIA neom agent toolkit uses Amazon s3 vectors for agent memory · 1 src
- OpenAI launches dots agent to handle ongoing tasks without human input · 2 src
- Anthropic deploys Claude buying agent to handle inbound sales · 2 src
- Amazon releases Strands Decider 2B, a Jev-inspired open-source decision model · 1 src
Comments
via GitHub Discussions