Uploaded September 2026 | Updated November 2026, 2 days ago
This video walks you through the guide at arize.com/docs/ax/cookbooks/evaluate/evaluate-receipt-agents-with-image-judge.
LLM-as-a-judge evaluation is not limited to text. You can also evaluate multimodal AI applications by using images as the source of truth.
In this Arize AX walkthrough, see how to evaluate a receipt extraction agent with a multimodal LLM-as-a-judge evaluator.
The example agent takes a receipt image and extracts structured JSON, including the merchant, line items, amounts, currency, subtotal, and total. The challenge is that checking the JSON format alone does not tell you whether the extracted values actually match the image.
To solve that, we build a visual groundedness evaluator that takes the original receipt image, independently interprets its contents, and compares them against the agent’s extracted output.
In this tutorial, you’ll learn how to:
• Trace a multimodal AI application in Arize AX
• Evaluate image inputs with LLM-as-a-judge
• Use the original image as ground truth for an evaluation
• Compare structured JSON extraction against receipt contents
• Build a visual groundedness evaluator
• Filter evaluations to specific LLM spans
• Map image and output fields into an evaluator prompt
• Define grounded, not grounded, and needs review outcomes
• Run evaluations continuously on production traces
• Sample production traffic to control evaluation cost
• Use a stronger judge model to validate a cheaper extraction model
• Surface cases that should be escalated to human review
The demo also shows why multimodal judges can catch more than simple extraction mismatches. In one example, the judge notices that the individual receipt items do not add up to the reported subtotal, flagging a consistency issue even though the extracted values themselves appear plausible.
This creates a practical evaluation pattern for document extraction, receipts, invoices, forms, screenshots, and other multimodal AI workflows.
Chapters:
00:00 LLM-as-a-judge for multimodal content
00:32 Evaluating a receipt extraction agent
01:22 Why text-only evals are not enough
02:03 Building a visual groundedness evaluator
03:06 Set up the example application
04:06 Trace receipt images in Arize AX
04:56 Create the image-based LLM judge
06:10 Configure labels and span filters
07:04 Map the image and extraction fields
07:45 Run the evaluator continuously
08:10 Review grounded and failed examples
08:48 Catching inconsistencies the agent missed
🔗 Explore Arize AX: arize.com/ax
🔗 Read the Arize docs: arize.com/docs/ax
🔔 Subscribe for more videos on LLM evals, multimodal AI, agent observability, and AI engineering:
youtube.com/@arizeai?sub_confirmation=1
#LLMEvals #MultimodalAI #LLMasAJudge
This video walks you through the guide at arize.com/docs/ax/cookbooks/evaluate/evaluate-receipt-agents-with-image-judge.
LLM-as-a-judge evaluation is not limited to text. You can also evaluate multimodal AI applications by using images as the source of truth.
In this Arize AX walkthrough, see how to evaluate a receipt extraction agent with a multimodal LLM-as-a-judge evaluator.
The example agent takes a receipt image and extracts structured JSON, including the merchant, line items, amounts, currency, subtotal, and total. The challenge is that checking the JSON format alone does not tell you whether the extracted values actually match the image.
To solve that, we build a visual groundedness evaluator that takes the original receipt image, independently interprets its contents, and compares them against the agent’s extracted output.
In this tutorial, you’ll learn how to:
• Trace a multimodal AI application in Arize AX
• Evaluate image inputs with LLM-as-a-judge
• Use the original image as ground truth for an evaluation
• Compare structured JSON extraction against receipt contents
• Build a visual groundedness evaluator
• Filter evaluations to specific LLM spans
• Map image and output fields into an evaluator prompt
• Define grounded, not grounded, and needs review outcomes
• Run evaluations continuously on production traces
• Sample production traffic to control evaluation cost
• Use a stronger judge model to validate a cheaper extraction model
• Surface cases that should be escalated to human review
The demo also shows why multimodal judges can catch more than simple extraction mismatches. In one example, the judge notices that the individual receipt items do not add up to the reported subtotal, flagging a consistency issue even though the extracted values themselves appear plausible.
This creates a practical evaluation pattern for document extraction, receipts, invoices, forms, screenshots, and other multimodal AI workflows.
Chapters:
00:00 LLM-as-a-judge for multimodal content
00:32 Evaluating a receipt extraction agent
01:22 Why text-only evals are not enough
02:03 Building a visual groundedness evaluator
03:06 Set up the example application
04:06 Trace receipt images in Arize AX
04:56 Create the image-based LLM judge
06:10 Configure labels and span filters
07:04 Map the image and extraction fields
07:45 Run the evaluator continuously
08:10 Review grounded and failed examples
08:48 Catching inconsistencies the agent missed
🔗 Explore Arize AX: arize.com/ax
🔗 Read the Arize docs: arize.com/docs/ax
🔔 Subscribe for more videos on LLM evals, multimodal AI, agent observability, and AI engineering:
youtube.com/@arizeai?sub_confirmation=1
#LLMEvals #MultimodalAI #LLMasAJudge










