Small parsing models are beating frontier models at reading the documents agents rely on
LlamaIndex's new OCRia-1.5 improves table and form scores, its own ranking finds small specialized models beating frontier models on structure, and Unstructured now turns audio and video into the same structured elements as text.
- OCRia-1.5 improved Agentic table scores on ParseBench-2 from 86.5 to 90.1, and Cost Effective visual grounding scores from 77.0 to 88.5, according to LlamaIndex. Self-reported m:…9579
- In LlamaIndex's ranking of OCR models, OCRia-1 scores 91.5 on ParseBench tables against 94.26 for Claude Opus 5.5, while AWS Textract and Google Document AI charge $1.50 per 1,000 pages for text reading. Self-reported m:…9578
- Schema drift changes the structure of a data source in ways downstream systems do not expect, and silent corruption happens when pipelines cast data types without raising errors; LlamaIndex recommends validating the structured output of extraction steps. Self-reported m:…9576
- Unstructured's Multimedia Partitioner turns audio and video into structured JSON elements, and a NASA video test case yielded 10 TranscriptFragment and 9 SceneDescription elements. Self-reported m:…5025
The small model that reads tables better
LlamaIndex has released OCRia-1.5, an upgraded document parsing model available in two sizes: the 4B flagship for agentic workflows and a new 2B model for cost-effective tasks. The smaller model has half the parameters of OCRia-1, and the update improves accuracy on complex tables, charts, and forms, while enhancing the precision of visual grounding bounding boxes. Both models are now integrated into LlamaParse, and the company reports new performance marks on ParseBench-2, its own benchmark for document parsing. The interesting part is not the size of the improvement but where it lands: the hard parts of documents, the tables and forms where a wrong cell changes the answer an agent gives. m:…9579
On ParseBench-2, Agentic table scores rose from 86.5 to 90.1, and Cost Effective table scores rose from 83.2 to 87.5. Agentic chart scores rose from 74.9 to 76.8, and Cost Effective chart scores rose from 67.3 to 73.0. The Agentic forms score rose by 2.6 points and the Cost Effective forms score by 4.2 points. Visual grounding, the ability to point at the right place on the page, improved most sharply at the cheaper tier, with the Cost Effective score rising from 77.0 to 88.5. These are LlamaIndex's own numbers on its own benchmark. The cheaper tier is catching up fastest, which matters because document parsing is a volume business where per-page cost decides whether a pipeline is affordable. m:…9579
Specialists beat generalists on structure
LlamaIndex's latest analysis of OCR models makes a broader claim: specialized, smaller models are increasingly outperforming general-purpose frontier models in document parsing at a fraction of the cost and latency. In its ranking, OCRia-1 scores 91.5 on the ParseBench table dimension, against 94.26 for Claude Opus 5.5, a much larger model. OCRia-1 is a 4-billion-parameter model that runs inside LlamaParse, with a median latency of 2.2 seconds per page. LlamaParse's Agentic Plus tier has the highest overall score of all 155 entries on the ParseBench leaderboard at 86.69. ParseBench covers about 2,000 human-verified pages pulled from insurance filings, financial reports, and government documents. m:…9578
The comparison with the large cloud services is about what you pay for and what you get. AWS Textract and Google Document AI charge $1.50 per 1,000 pages for text reading, and LlamaIndex's analysis says they remain cost-effective for plain text but lag in structural extraction tasks like nested headers and chart data. Azure Document Intelligence supports roughly 300 printed languages, and Mistral OCR 4.1 has a hosted latency of 1.8 seconds at the median. The decision for a developer is therefore not which model is smartest in general, but which one reads the specific structure of their documents correctly, at a price and latency the pipeline can sustain. m:…9578
The failure mode nobody sees, and the data beyond documents
Better parsing does not remove the failure mode that matters most, because the failure is silent. Schema drift occurs when the structure of a data source changes in a way that diverges from what the downstream system expects, and it is distinct from data quality issues, which affect values inside a schema rather than the structure itself. Column removal and data type changes are the most damaging forms of schema drift, and value range drift is particularly problematic for document-heavy workflows where new categorical values can cause misbehavior in classifiers. Silent corruption occurs when pipelines cast data types without raising errors, allowing incorrect data to flow downstream. The pipeline keeps running, the agent keeps answering, and nobody knows the numbers changed. m:…9576
The defense LlamaIndex recommends is layered: structural comparison at ingestion prevents corrupted data from entering the pipeline by checking for expected columns and data types, statistical profiling catches value range drift and subtle type changes that structural comparison misses, and a schema registry provides a centralized store for tracking structural changes and triggering downstream notifications. Critically, schema validation should be applied to the structured output of document extraction steps, to catch cases where source document changes cause empty or incorrect fields. LlamaParse handles document layout changes inherently, according to the company, but still requires output schema validation to ensure extracted data matches downstream expectations. Continuous integration pipelines can also catch schema changes by validating against reference documents or stored expected schemas. m:…9576
Meanwhile, the definition of what counts as a document is widening. Unstructured has launched a Multimedia Partitioner designed to transform audio and video files into structured JSON elements for AI applications, creating time-stamped TranscriptFragment and SceneDescription elements so agents can access multimedia context without processing opaque media files. The system uses Parakeet TDT 0.6B v3 for speech transcription, pyannote Community-1 for speaker diarization, and WeSpeaker to match voices across audio windows for consistent speaker labeling, sampling video frames within 300-second windows to extract visual metadata. The architecture supports OpenAI-compatible inference backends for transcription and scene descriptions. In a NASA video test case, the system yielded 10 TranscriptFragment and 9 SceneDescription elements. m:…5025
The thread connecting these releases is that agents are moving from reading text to consuming structured data, and the quality of that data now decides the quality of the answer. A better table parser, a validated schema, and a transcript that arrives as structured elements are all the same project: give the agent inputs it can trust, and check them before they reach it. m:…9579 m:…9576 m:…5025
What is still unresolved
- Whether independent testing confirms the OCRia-1.5 gains on documents outside ParseBench's insurance, financial, and government pages.
- How many production pipelines validate extracted output against a schema today, and what share of agent errors trace back to silent drift.
- Whether multimedia partitioning reaches production use beyond test cases like the NASA video, and how accurate its transcripts and scene descriptions are at scale.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.