Skip to main content

Search

Let us guide you!

Configure the “Private and Secure AI Action”

Updated on

The Ask AI action with the Private & Secure AI source lets you run open-source AI models hosted on TimeTonic's sovereign infrastructure in France.

Sovereign data

Processed data never leaves the environment controlled by TimeTonic. No transfer to any external provider.

TimeTonic credits

Each AI call consumes TimeTonic credits (1 credit = 1,000 tokens). Consumption varies depending on the selected model.

Open-source models

Mistral, LLaMA, Qwen, and other models deployed on TimeTonic's infrastructure.

1. Model selection

Choose the AI model to use from those available on TimeTonic's sovereign infrastructure. Each model displays its maximum context window and credit cost directly in the list.

Model Total
context Input
/M Output
/M Ideal for
Mistral 7B Instruct
Efficient
32.7K 100K 100K Repetitive text, simple generation, high-volume tasks, sensitive data
Llama 3.1 8B Instruct
Efficient
131K 100K 100K Basic extraction, fast classification, sovereign data triage
GPT-OSS 20B
Efficient
131K 40K 150K Lightweight reasoning, lead qualification, short analysis on internal data
Mistral Small 3.2 24B
Efficient 📎 Vision 📄 OCR
128K 90K 280K Native French support, simple image analysis, text extraction, sensitive data
Llama 3.3 70B Instruct
Premium
131K 670K 670K Complex text documents, legal deeds, contracts, reports on sensitive data
Qwen 2.5 VL 72B
Premium 📎 Vision 📄 OCR
32.7K 910K 910K OCR specialist: precise extraction from scanned documents, ID cards, invoices, bank details

📎 Vision: the model can analyze an attached image and understand its content
📄 OCR: the model can extract structured text from document photos
Total context = combined input + output · Input /M and Output /M = credits per million tokens · K = thousands · M = millions

2. AI role description

This field defines the expected behavior and context for the model. A precise role significantly improves the quality and consistency of responses.

"You are an assistant responsible for extracting structured information."

"You are an expert in analyzing administrative documents."

3. Prompt

The prompt is the precise instruction sent to the model. Use the $ symbol to dynamically inject values from your fields.

💡 The more precise and contextualized the prompt, the more relevant the response will be.

For PDF or image analysis, a specific prompt is required → See OCR PDF & images prompts

4. Attachments

Select a field containing image files to allow the AI to analyze their content. Useful for:

📄 Document analysis
🔍 Information extraction from files
🖼️ Image processing
⚠️ Only models with the 📎 Attachments badge support this option. Verify compatibility before using in production.

For PDF file processing → See the OCR PDF & images procedure

5. Maximum token count

The Maximum token count field caps the length of the response generated by the model.

Token limitEffect on responseUse caseRecommended value
Low
Short, concise response: the AI gets straight to the pointClassification, single-value extraction, category, yes/no50 to 200 tokens
Medium
Structured response in a few linesSuggested email reply, short reformulation100 to 400 tokens
High
Long, detailed response: summary, report, full analysisDocument summary, meeting notes, content generation500 to 1,000 tokens
Adjust this parameter to match the nature of the task → limiting tokens helps control your TimeTonic credit consumption. A 100-token response costs ten times less than a 1,000-token response.

The token count directly impacts your credit consumption. To anticipate the impact before configuring your action → Estimate your AI credit consumption

6. Response format

The response format determines how the AI-generated response will be structured and stored in your notebook. Two options are available:

Text

The response is stored as-is in a text field. Ideal for summaries, comments, or free-form responses.

Structured JSON

Each JSON property is stored in a dedicated field. Recommended for structured extractions and automated processing.

💡 The Structured JSON format requires a specific prompt in the Prompt field to instruct the model on which data to extract and in what structure → especially for PDF or image analysis. Each JSON key then maps to a destination field in your record.

See the OCR PDF & images configuration →

7. Response field

This field determines where the AI response will be stored in your notebook. Its behavior depends directly on the format selected in the previous step.

If you selected the Text format

The Response field appears. Select the text field in your notebook where the generated response will be stored. Only one destination field is required.

If you selected the Structured JSON format

The Response field disappears and is replaced by a mapping table. For each JSON property you requested in your prompt, you must:

  1. Enter the exact name of the JSON property as you defined it in your prompt (e.g., name, date_of_birth, amount)
  2. Select the corresponding destination field in your notebook where the extracted value will be stored

Repeat for each property to extract. You can add as many rows as needed using the + button.

Example: You are extracting data from an ID card and your prompt requests a JSON with the keys last_name, first_name, and date_of_birth. You create three rows and map each key to the corresponding field in your notebook.

8. Credit consumption

Learn more

AI credits by license plan

Read the article →

Learn more

Consumption in the logs

Read the article →

Learn more

Credits exhausted: what to do?

Read the article →

Best practices

  • Define a precise role before writing the prompt.
  • Limit the token count when a short response is sufficient.
  • Test the automation with realistic data.
  • Check consumption in the logs after the first runs.