# Image Recognition

How to get the best results when your AVCodex agent analyzes images. Useful when techs send phone photos of error screens, racks, terminations, or signal flow diagrams.

---

Your AVCodex agent can analyze images uploaded by users. This guide explains how image recognition works and how to dial it in for AV work.

## How it works

When a user uploads an image, AVCodex processes it in one of two ways depending on the agent's model.

### Vision-capable models (recommended)

Models with native vision see images directly, the way a person would. The image is embedded in the conversation and the model can reference it naturally.

**Vision-capable models:**

| Provider | Models |
|---|---|
| OpenAI | GPT-5, GPT-5 Mini, GPT-5 Nano, GPT-4.1, GPT-4.1 Mini, GPT-4.1 Nano |
| Anthropic | Claude Opus 4.1, Claude Opus 4, Claude Sonnet 4.5, Claude Sonnet 4, Claude 3.7 Sonnet, Claude 3.5 Haiku |
| Google | Gemini 2.5 Pro, Gemini 2.5 Flash, Gemini 2.5 Flash Lite, Gemini 2.0 Flash, Gemini 2.0 Flash Lite |

**Models WITHOUT vision:**

- OpenAI o-series (o1, o1 Pro, o3, o3 Pro, o3-mini, o4 Mini).

### Non-vision models (fallback)

For models without native vision (the o-series reasoning models), AVCodex uses a separate image analysis tool powered by OpenAI's GPT-5 to describe the image, then passes that description to your agent's model.

That two-step process can lose context compared to native vision models. For most AV use cases (a tech pointing a phone camera at a screen full of small text and tiny LEDs), native vision is the better path.

## Getting the best results

### 1. Pick the right model

For agents that lean on image analysis (rack photos, error code screens, drawings, label reading), pick a vision-capable model.

1. Open your agent in the AVCodex dashboard.
2. Go to **Build** > **Configure**.
3. Under **Model**, choose a vision-capable model like GPT-5 or Claude Sonnet 4.

### 2. Enable image recognition

Make sure the capability is on:

1. Go to **Build** > **Actions**.
2. Find **Image Recognition** in the Basic Actions list.
3. Toggle it **ON**.

### 3. Image quality tips

For best recognition accuracy:

- **Resolution**: higher-resolution images produce better results. A clear shot of a touch panel beats a tiny crop.
- **Lighting**: well-lit images are easier to analyze. Closet lighting can be rough on a phone camera, so flashlight on if needed.
- **Focus**: keep the subject in focus. Blurry SIMPL error text is hard for any model.
- **Format**: JPEG, PNG, and WebP are all supported.

### 4. Prompting for analysis

Guide users to be specific about what they want analyzed.

**Good prompts:**

- "What error code is shown on this DM-NVX touch panel?"
- "Identify the model and serial of this Crestron rack-mount unit."
- "Read the text on this Q-SYS Designer fault dialog."
- "What signals are flowing into the DSP in this drawing?"

**Vague prompts:**

- "What is this?"
- "Tell me about the image."

## Comparing to ChatGPT

If you notice differences between your AVCodex agent and ChatGPT's native image analysis:

1. **Model selection**: ChatGPT uses GPT-5 by default. Make sure your AVCodex agent uses GPT-5 or a comparable vision model like Claude Sonnet 4.
2. **System prompt**: your agent's role and instructions affect how it interprets images. ChatGPT has different defaults.
3. **Context window**: ChatGPT handles conversation context differently. For complex multi-image analysis (multiple rack angles, multi-page drawings), results can vary.

## Troubleshooting

### Images not being analyzed

- Verify **Image Recognition** is enabled in the agent's Actions.
- Check that the user is uploading supported formats (JPEG, PNG, WebP, GIF).
- Make sure images are from supported sources (direct uploads, not every external URL).

### Poor accuracy

- Switch to a vision-capable model (GPT-5, Claude Sonnet 4, etc.).
- Ask users to provide clearer, higher-resolution images. Encourage close-ups for label and serial reads.
- Add specific instructions in your system prompt about how to analyze images (e.g., "When a user sends a rack photo, identify each device, manufacturer, and rack unit position from top to bottom.").

### Slow response times

Image analysis typically takes 5 to 15 seconds depending on complexity. For faster responses:

- Use smaller image files when possible.
- Consider GPT-5 Nano or Claude 3.5 Haiku for the speed-vs-accuracy trade-off.

## Model comparison for vision

| Model | Vision quality | Speed | Best for |
|---|---|---|---|
| GPT-5 | Excellent | Medium | Complex reasoning about drawings, signal flow, multi-element rack photos. |
| Claude Sonnet 4.5 | Excellent | Fast | Detailed visual descriptions of racks and labeled gear. |
| Gemini 2.5 Pro | Excellent | Medium | Multi-image comparisons, long context (multi-page drawings). |
| GPT-5 Mini | Very good | Fast | Balanced performance for routine field photos. |
| GPT-5 Nano | Good | Very fast | Quick, simple analysis (a touch panel error code lookup). |
| Claude 3.5 Haiku | Good | Very fast | Fast responses for high-volume room support agents. |
| Gemini 2.5 Flash | Very good | Fast | Balanced speed and quality. |

*AVCodex · Your AV expertise. Amplified by AI.*