Skip to main content
Every Safe has two instruction fields that give you control over the AI’s behavior at different stages: ingestionInstructions and instructions.
  • ingestionInstructions — evaluated at graph build time, when content is added to the Safe. Controls what the knowledge graph captures.
  • instructions — evaluated at retrieval and agent time, when the AI answers a question. Controls how answers are formed.

Ingestion instructions

Guide graph construction: what entities to extract, what to skip, and how to weight relationships.

Retrieval instructions

Shape AI answers: tone, scope constraints, source prioritization, and behavioral rules.
Both fields accept plain text up to 100,000 characters. You can update them at any time from the Along dashboard or via the API — no re-deployment required.

​
How they work together

Ingestion shapes the graph structure; retrieval instructions shape how the graph is queried and how answers are composed. Think of ingestion as “what to remember” and retrieval as “how to think.” When a document is added to a Safe, Along runs extraction using ingestionInstructions as guidance. The resulting knowledge graph reflects those choices — which entities were captured, which relationships were created, what was skipped. Later, when an AI assistant connects to the Safe via MCP and asks a question, instructions travel with every response, telling the AI how to reason over and present what the graph contains. Neither field depends on the other. You can use one, both, or neither. But used together, they give you precise control over the full lifecycle: from raw content into structured knowledge, and from structured knowledge into useful answers.