NVIDIA adds open AI models to help agents search charts, tables and text

AI agents searching business documents have new open models for finding evidence in charts, tables and text—not just processing an answer afterward. NVIDIA released its Nemotron 3.5 Embed, Rerank and Parse collection on October 21, saying better retrieval reduced agent inference-token use while improving retrieval accuracy in its tests.

The release covers three connected jobs: making document collections searchable, moving the strongest matches up the results, and turning document images into structured text. NVIDIA describes the models as open and commercially available, with embedding options at two sizes and a separate, smaller parser.

Finding a page is not the same as finding evidence

An embedding model represents queries and documents as numerical vectors so their similarity can be compared. Nemotron 3.5 Embed extends NVIDIA’s earlier text models to document images, including charts, infographics, datasheets and schematics. That lets applications search visual information alongside written content.

The reranker takes a second look. Rather than encoding a query and document separately, it evaluates each pair together and assigns a relevance score. Its job is to distinguish a broadly related result from evidence that directly answers the question before an agent uses it.

The released models

• Embed: eight-billion- and one-billion-parameter models in BF16, plus a one-billion-parameter NVFP4 version. These are different numerical formats for the model weights; NVIDIA positions NVFP4 for higher-throughput retrieval on supported hardware.
• Rerank: an eight-billion-parameter BF16 model that reorders retrieved candidates across text and image inputs.
• Parse: a 631-million-parameter model that converts document images into structured text, with layout classes, bounding boxes locating content, and reading order.

The token savings come from changing retrieval

NVIDIA tested retrieval configurations across ViDoRe v3, BRIGHT and BrowseComp-Plus. It held the agent’s language model, tools, tasks, document collection and execution policy fixed. The agent could search, inspect evidence and refine its next search; the experiments changed how that search tool retrieved and ranked results.

Adding Rerank cut agent inference-token use per query by approximately 14% with Embed 8B and 13% with Embed 1B, while improving retrieval accuracy. NVIDIA counted input and output tokens across all agent turns. Those percentages describe agent token use under the tested configurations, rather than an across-the-board reduction in operating costs.

The company also reports a 61.17 score for Embed 8B on ViDoRe v3.1, first among dense retrieval models as of October 20. Rerank scored 60.49 against Qwen-3-VL-Reranker-8B’s 59.17. Both use nDCG@10, a measure of how well relevant results appear near the top of a ranked list.

Existing indexes and custom documents

For existing users, NVIDIA says the embedding models’ language component stayed frozen during training. Text-only embeddings are therefore expected to remain unchanged from the corresponding Nemotron 3 Embed models, allowing existing text embeddings and vector indexes to be reused.

Teams can also adapt retrieval to their own documents using NVIDIA’s fine-tuning recipe. In a chart-and-infographic experiment evaluated on held-out synthetic queries, the company reports relative retrieval-score gains of 18% for Embed 1B and 3% for Embed 8B. These are domain-adaptation results, not gains promised for every collection.

Parsing has its own accuracy profile. NVIDIA’s Parse model card lists a ParseBench mean score of 75.20, with 84.67 for tables and 64.4 for charts. Developers can access the collection’s weights on Hugging Face and deploy through NVIDIA NIM or supported inference partners.