Retrieval Instructions

Shape how the AI answers questions: tone, scope, context prioritization, and behavioral rules.

instructions are sent to the AI assistant at the start of every MCP session and prepended to every response from the along tool. They define the AI’s persona, scope constraints, answer style, and behavioral rules when working with a Safe’s knowledge.

Unlike ingestionInstructions, which run once when content is added, retrieval instructions are active throughout every AI interaction — they travel with the data.

When to use retrieval instructions

Use instructions when you need the AI to:

• Stay focused on a specific topic area and decline out-of-scope questions
• Maintain a particular tone or persona (formal, technical, accessible)
• Always cite sources or use a specific citation format
• Follow formatting conventions (bullet points, structured summaries, prose)
• Apply hard behavioral rules (never speculate, always ask for clarification when ambiguous)

Parameters

Field	Value
Type	string
Max length	100,000 characters
Evaluated	At session initialization (sent as MCP server instructions) and prepended to every along tool response
Default	null — no custom instructions

What you can control

• Tone and persona — define whether the AI is formal or informal, expert-facing or accessible
• Scope constraints — restrict answers to a specific topic domain
• Context prioritization — prefer recent documents, primary sources, or specific content types
• Answer format — bullet points, prose, structured templates, citation style
• Behavioral rules — hard constraints like “never speculate” or “always ask for clarification”

How instructions are delivered

When a user connects to a Safe via MCP, Along sends the instructions field as MCP server instructions during the initialize handshake. This means the AI assistant receives the instructions before any tool is called.

Every subsequent response from the along tool also prepends the instructions as a context header — [Instructions: ...] — at the start of the tool’s output. This ensures the instructions remain in scope for the AI throughout the session, even in long conversations where early context might otherwise be pushed out.

Examples

Customer success team

You are a customer success assistant for Acme Corp. Only answer questions about customer accounts, product usage, and support history in this Safe.

Always cite the source document when providing information (format: [Document Title, Date]).

If you cannot find a definitive answer in the knowledge base, say so clearly rather than speculating. Do not infer account status or customer sentiment beyond what is explicitly documented.

Sales team

This Safe contains sales intelligence. Prioritize recent deal notes and CRM data over older documents.

When discussing competitors, present only factual information sourced from recorded calls or documents — do not infer competitive positioning.

Format all deal summaries with these fields:
- Stage
- Key contacts
- Last activity
- Next steps

If a field is not documented, write "Not recorded" rather than omitting it.

Engineering knowledge base

Answer technical questions concisely. Prefer precision over completeness — a shorter accurate answer beats a longer uncertain one.

When referencing architecture decisions, always note the date of the decision document.

Flag any information older than 12 months as potentially outdated: append "(Note: sourced from [date] — verify current state)" to the relevant claim.

Do not answer questions outside the scope of this knowledge base. If asked about something not covered, say "That's not documented in this Safe."

Safe API keys can also carry their own instructions field. When a request is made using a Safe API key that has its own instructions, those instructions are appended after the Safe’s instructions — separated by a horizontal rule. This lets different teams or integrations use the same Safe with different behavioral contexts, without needing to create separate Safes.