> ## Documentation Index
> Fetch the complete documentation index at: https://explore.airia.com/llms.txt
> Use this file to discover all available pages before exploring further.
# AI Model Context Settings
> Control what the model knows at runtime — date/time, user identity, conversation history, attachments, and user input.
Every AI Model step in Agent Studio has a **Context** panel that controls which runtime information is sent alongside your prompt. These settings determine what the model can see beyond your instructions — the current time, who is asking, what files are attached, and how much of the conversation it remembers.
Getting context right has a direct impact on answer quality, token cost, and whether the model behaves consistently across different trigger sources.
***
## Quick reference
| Setting | Default | What it does |
| ----------------------------- | ----------- | ------------------------------------------------------------------------------------------------------------------- |
| **Date and time context** | On | Adds a system message with the current date and time |
| **Timezone** | UTC | Controls which timezone the date/time is presented in |
| **User details context** | On | Adds a system message with the user's profile and preferences |
| **Always include user input** | On | Ensures the original user message reaches the model, even when the step is not directly connected to the Input step |
| **Include Attachments** | On | Passes uploaded files and images into the model's context |
| **Include the chat history** | All history | Sends previous conversation turns so the model can follow multi-turn dialogue |
***
## Date and time context
When enabled, Airia injects a system message before your prompt:
```
The current date and time is 2026-03-26T14:30:00+00:00;
please use this information for any references to the present moment.
```
The timestamp is generated fresh on every execution and formatted in ISO 8601. The model uses it to ground any time-sensitive reasoning — deadlines, scheduling, relative dates ("last week"), or seasonal context.
* Agents that answer questions about dates, deadlines, or schedules
* Workflow agents triggered on a cron that need to know "today"
* Any prompt that includes phrases like "as of today" or "this quarter"
* Pure classification tasks (e.g., sentiment analysis, intent detection) where the date is irrelevant
* Translation or formatting agents that process text without time context
* When you want to save a small number of tokens on high-volume, time-insensitive pipelines
Even with this setting enabled, you can also reference the same value inside your prompt text using the `{{ Helpers.CurrentDateTime }}` expression. The context setting provides it as a separate system message; the expression lets you embed it inline. You can use both — they complement each other.
***
## Timezone
Controls the timezone applied to the date/time context message. The dropdown offers three modes:
| Mode | Behaviour | Best for |
| ----------------- | ---------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------ |
| **UTC** (default) | Timestamp is always in UTC (`+00:00`) | Backend agents, cron jobs, internal tooling where a consistent reference is preferred |
| **Custom** | A fixed IANA timezone you choose (e.g., `America/New_York`, `Europe/London`) | Agents that serve a known geography — e.g., a customer support bot for US East Coast hours |
| **Client** | Uses the end user's browser timezone, sent at execution time | Global-facing chat agents where each user expects times in their own local zone |
When a custom or client timezone is set, the injected message includes the timezone name:
```
The current date and time is 2026-03-26T10:30:00-04:00 (Eastern Daylight Time);
please use this information for any references to the present moment.
```
If the client timezone is unavailable at runtime (e.g., the agent was triggered via API without timezone data), Airia falls back to UTC silently.
***
## User details context
When enabled, Airia injects a system message containing the authenticated user's profile:
```
Consider the following preferences and information about the user
when crafting your response:
{"General":"Senior account manager at Acme Corp","WorkDescription":"Manages enterprise renewals","PersonalPreferences":"Prefers concise bullet points over long paragraphs"}
```
This includes three fields from the user's Airia profile:
| Field | What it contains |
| --------------------- | -------------------------------------------------------- |
| `General` | Free-text bio or role description |
| `WorkDescription` | The user's job function or responsibilities |
| `PersonalPreferences` | Communication style, formatting, or language preferences |
* Personalised assistants that adapt tone, detail level, or vocabulary to the user
* Agents that need to know the user's role to scope their answers (e.g., executive summary vs. technical deep-dive)
* Internal tools where users have filled in their Airia profile
* Public-facing or anonymous agents where no user profile exists
* Agents processing data in bulk (e.g., document classification pipelines) where personalisation adds noise
* Privacy-sensitive workflows where user details should not be sent to the model provider
User profile data is separate from the `{{ User.FirstName }}`, `{{ User.Email }}` template expressions. The context setting sends the full profile JSON as a system message. The expressions let you embed individual fields inline in your prompt. See the [Prompts](/building-and-deploying-agents/agent-basics/prompts#user) page for the full list of `User` properties.
***
## Always include user input
In a multi-step agent, not every AI Model step is directly connected to the Input step. For example, a three-step pipeline might look like:
```
Input → Step 1 (Classifier) → Step 2 (Responder)
```
Step 2 receives its input from Step 1, not from Input. Without this setting, Step 2 would only see the classifier's output — it would have no idea what the user originally asked.
When **Always include user input** is enabled (the default), Airia ensures the original user message is added to the model's context regardless of graph wiring.
* Most agents — you almost always want the model to see the user's original question
* Multi-step agents where downstream steps need the original request for grounding
* Agents where the upstream step transforms or summarises the input, but the final model still needs the raw question
* A dedicated summariser step that should only process the output of a previous step, not the raw user query
* Steps that intentionally isolate the model from the user's wording (e.g., a reformulation step whose output replaces the original)
* Agents triggered by webhooks or schedules where `UserInput` is empty or a system payload — including it would add noise
Disabling this on a step that is the only AI Model in your agent usually means the model receives no user message at all. Only disable it when the step genuinely should not see the original input.
***
## Include Attachments
When enabled, any files or images the user uploads with their message are added to the model's context. This includes:
| Included data | Description |
| ------------- | --------------------------------------------------- |
| File name | The original filename (e.g., `Q1-report.pdf`) |
| Content type | MIME type (e.g., `application/pdf`, `image/png`) |
| File content | The file's content, processed for the model to read |
* Document Q\&A agents ("summarise this PDF", "extract the table from this image")
* Agents that analyse uploaded images (charts, screenshots, photos)
* Any agent where users are expected to attach files as part of their workflow
* Agents that never receive file uploads — disabling saves context window space
* Text-only chat assistants where attachments would be irrelevant
* High-volume pipelines where accidentally large attachments could inflate token costs
Not all models support all file types. Vision-capable models (GPT-4o, Claude 3.5 Sonnet, Gemini Pro Vision, etc.) can process images natively. For non-vision models, only text-extractable content (like parsed PDFs) will be useful. Check your model's capabilities before relying on attachment processing.
***
## Include the chat history
Controls how much of the conversation the model can see. Click the dropdown to choose one of three modes:
Every previous message in the conversation is included. The model has full context of everything said so far.
Only the most recent N messages are included. You specify the number. Older messages are silently dropped.
No previous messages are included. The model treats every request as if it is the first message in the conversation.
### Choosing the right mode
| Mode | Best for | Trade-off |
| ------------------- | --------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------- |
| **All history** | Conversational assistants, multi-turn reasoning, coaching bots | Token usage grows with conversation length. Long conversations can hit context window limits. |
| **Last N messages** | Assistants with long sessions where only recent context matters (e.g., customer support, troubleshooting) | Good balance of continuity and cost control. Experiment with N — 10–20 messages is a common starting point. |
| **No history** | Single-shot tasks (classification, data extraction), webhook-triggered agents, scheduled pipelines | No conversational memory. Each execution is stateless. Lowest token cost. |
If your agent uses the **Debug** tab and you notice the model is repeating itself or contradicting earlier answers, check your chat history setting. "All history" with very long conversations can push important context out of the model's effective attention window. Switching to "Last N" often fixes this.
***
## How context settings relate to prompt expressions
Context settings and [prompt expressions](/building-and-deploying-agents/agent-basics/prompts) are two complementary ways to give the model runtime information. They are **not** mutually exclusive — many agents use both.
| Mechanism | How it works | When to use |
| -------------------------------- | ------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------- |
| **Context settings** (this page) | Airia injects pre-formatted **system messages** automatically. You do not write them — they appear alongside your prompt. | When you want the information available to the model without embedding it in your prompt text. Simple toggle, zero prompt editing. |
| **Prompt expressions** (`{{ }}`) | You write Scriban expressions directly inside your prompt. They are evaluated at runtime and replaced with values. | When you need precise control over wording, placement, or conditional logic around the data. |
**Example — Date/time both ways:**
The **Date and time context** toggle adds this system message automatically:
```
The current date and time is 2026-03-26T14:30:00+00:00;
please use this information for any references to the present moment.
```
The **prompt expression** `{{ Helpers.CurrentDateTime }}` lets you embed the same timestamp inside your own instructions:
```
Today is {{ Helpers.CurrentDateTime }}. Any deadline the user mentions
should be compared against this date.
```
Using both is fine — the model sees the information twice, but from different angles. The context setting provides a generic grounding message; the expression lets you frame it specifically for your use case.
***
## Recommended presets by agent type
A general-purpose chat agent that talks to end users.
| Setting | Value | Why |
| ------------------------- | --------------- | --------------------------------------------- |
| Date and time context | **On** | Users ask time-sensitive questions |
| Timezone | **Client** | Each user sees times in their own zone |
| User details context | **On** | Personalises tone and depth |
| Always include user input | **On** | Model always sees the question |
| Include Attachments | **On** | Users may share files |
| Include the chat history | **All history** | Multi-turn conversation requires full context |
An agent triggered by API or webhook that classifies incoming text and returns a category.
| Setting | Value | Why |
| ------------------------- | -------------- | ----------------------------------------------- |
| Date and time context | **Off** | Classification is time-independent |
| Timezone | — | N/A (date/time is off) |
| User details context | **Off** | Classification should not depend on who sent it |
| Always include user input | **On** | The input is the text to classify |
| Include Attachments | **Off** | Text-only input |
| Include the chat history | **No history** | Each call is independent |
A multi-step agent that receives a file, extracts data, and produces a structured output.
| Setting | Value | Why |
| ------------------------- | -------------- | ------------------------------------------------------------------------ |
| Date and time context | **On** | Useful if documents reference relative dates |
| Timezone | **UTC** | Consistent timestamp for audit trails |
| User details context | **Off** | Processing is content-driven, not user-driven |
| Always include user input | **Off** | The model should focus on the extracted document, not the upload message |
| Include Attachments | **On** | The file is the primary input |
| Include the chat history | **No history** | Each document is processed independently |
A cron-triggered agent that runs overnight to generate reports or process queues.
| Setting | Value | Why |
| ------------------------- | ------------------------------------- | -------------------------------------------------- |
| Date and time context | **On** | Report needs to know the current date |
| Timezone | **Custom** (e.g., `America/New_York`) | Reports are for a specific region |
| User details context | **Off** | No interactive user |
| Always include user input | **Off** | Input comes from the trigger payload, not a person |
| Include Attachments | **Off** | No file uploads in scheduled runs |
| Include the chat history | **No history** | Stateless execution |
***
## Troubleshooting
Enable **Date and time context** in the Context panel. Without it, the model can only guess the date based on its training data — which will be wrong.
If it is already enabled and the model still gets the date wrong, check whether your prompt contains conflicting date references (e.g., a hardcoded "Today is January 1, 2025" in a shared segment). The model may trust the explicit prompt text over the injected system message.
This usually happens with **All history** on long conversations. As the conversation grows, the model's effective attention can drift. Try switching to **Last N messages** (start with 10–20) to keep the context focused on the recent exchange.
Two common causes:
* **Chat history set to All**: Every previous message is sent on every turn. A 50-message conversation means 50+ messages in the context window on the 51st turn.
* **Attachments enabled with large files**: Uploaded documents consume significant tokens. If your agent does not process files, disable **Include Attachments**.
Use the **Debug** tab on any execution to inspect the full context sent to the model and identify which component is consuming the most tokens.
Check that **Always include user input** is enabled on the AI Model step that needs to see the original message. In multi-step flows, downstream steps do not receive the user's input automatically unless this setting is on.