Prompt Types

Every cached request is classified into one of seven prompt types. The type is stored in the usage log and surfaced in the dashboard's "Content types" strip and the prompts page filters.

---

Types

Type	Emoji	Color	What it means
text	📝	gray	Plain text messages — no special content
structured	⚙️	cyan	Explicit JSON output schema (response_format, responseSchema)
image	🖼	blue	Vision input — image in the prompt
audio	🎵	pink	Audio input — audio inline data or data URI
file	📄	orange	Document input — PDF, file reference, or non-image/audio data URI
tool	🔧	violet	Tool/function calling — tools, functions, or tool_choice present
image_output	🎨	rose	Image generation — eco.images.generate() or eco.images.edit()

---

Detection logic

Prompt type is detected from the raw request params before the cache key is hashed.

Priority order: structured > tool > file > structured (Gemini) > audio > image > text

structured

Triggered by response_format being present (any value) — OpenAI JSON mode, JSON schema, etc.

# OpenAI json_schema
{"response_format": {"type": "json_schema", "json_schema": {...}}}

# OpenAI json_object
{"response_format": {"type": "json_object"}}

Also triggered by Gemini's responseSchema or responseMimeType in generationConfig (detected via JSON scan after the tool check).

tool

Triggered when tools, functions, or tool_choice is present. This check runs before JSON serialization because Gemini's tool params may contain non-JSON-serializable proto objects.

{"tools": [{"type": "function", "function": {"name": "search"}}]}
{"functions": [...]}  # legacy OpenAI format
{"tool_choice": "auto"}

file

Triggered by:

• Anthropic document block: {"type": "document", ...}
• OpenAI Responses API file input: {"type": "input_file", ...}
• OpenAI file fields: "file_data" or "file_url" key present
• A data URI that is not image/audio/video: data:application/pdf;base64,...

audio

Triggered by:

• Gemini inline audio: "inlineData" present and "audio/" in the JSON
• Audio data URI: data:audio/...

# Gemini audio
{"inlineData": {"mimeType": "audio/wav", "data": "..."}}

# Data URI
{"content": "data:audio/mp3;base64,..."}

image

Triggered by:

• OpenAI vision: {"type": "image_url", ...}
• Anthropic image block: {"type": "image", ...}
• Gemini inline image: "inlineData" present and "image/" in the JSON
• Gemini Cloud Storage file: "fileData" key present
• Image data URI: data:image/...

Note: inlineData is only matched as image when the MIME type is an image type. Audio/video inlineData is handled by the audio check above.

image_output

Set directly by eco.images.generate() and eco.images.edit() — not detected from params. These methods tag every call (hit or miss) as image_output.

text

Everything else — a plain text conversation with no special content.

---

Dashboard display

The "Content types" strip on the overview page shows counts for any type that has at least one entry. Types with zero entries are hidden.

The prompts page filter bar includes buttons for all seven types so you can drill into any category.

---

Examples

// JS — text
{ model: 'gpt-4o', messages: [{ role: 'user', content: 'Hello' }] }
// → 'text'

// JS — structured
{ model: 'gpt-4o', messages: [...], response_format: { type: 'json_object' } }
// → 'structured'

// JS — tool
{ model: 'gpt-4o', messages: [...], tools: [{ type: 'function', function: { name: 'search' } }] }
// → 'tool'

// JS — image
{ model: 'gpt-4o', messages: [{ role: 'user', content: [{ type: 'image_url', image_url: { url: '...' } }] }] }
// → 'image'

// JS — audio (Gemini)
{ contents: [{ parts: [{ inlineData: { mimeType: 'audio/wav', data: '...' } }] }] }
// → 'audio'

// JS — file
{ messages: [{ role: 'user', content: [{ type: 'document', source: { type: 'base64', media_type: 'application/pdf' } }] }] }
// → 'file'