ingestionInstructions are evaluated when documents, transcripts, emails, and other content are added to the Safe. Along uses them to decide what entities to extract, what relationships to create, and what content to prioritize or skip.
Without ingestion instructions, Along applies its default extraction behavior across all content types. With them, you steer the graph toward the entities and relationships that matter for your use case.
When to use ingestion instructions
Use ingestionInstructions when you want to:
- Focus the knowledge graph on a specific domain (sales, legal, support, engineering)
- Exclude irrelevant content types (internal HR documents, boilerplate signatures, financial tables)
- Weight certain relationship types more heavily (always link people to their companies, always link issues to resolutions)
- Reduce graph noise by preventing extraction of ambiguous or low-signal entities
Parameters
| Field | Value |
|---|
| Type | string |
| Max length | 100,000 characters |
| Evaluated | At graph build time, per-document during ingestion |
| Default | null — Along uses its default extraction behavior |
What you can control
- Entity types to extract — specify exactly which types matter: people, companies, deals, topics, products, clauses, error codes, etc.
- Content to prioritize or skip — tell Along to focus on certain source types and ignore others
- Ambiguous or duplicate entities — give guidance on how to handle unclear or repeated references
- Relationship density — instruct Along to create more or fewer connections between entity types
Examples
Sales team
Extract: Person (name, title, company), Organization (name, industry), Deal (name, stage, value), Competitor (name).
Create WORKS_AT relationships between every Person and their Organization.
Create MENTIONED_IN relationships between Competitor and the Deal where they appear.
Skip: internal meeting notes that have no external attendees, personal calendar events, HR documents.
Link every person mention to their company using email domain when explicit company is not stated.
Legal team
Extract: Party (name, role), Clause (type, obligation), Obligation (description, due date), Document (title, execution date).
Create BOUND_BY relationships between Party and Clause.
Create OBLIGATED_TO relationships between Party and Obligation only when explicitly stated in the text.
Skip: financial tables, exhibits with only numerical data, cover pages.
Do not create relationships between Clause and individual people unless the clause explicitly names an individual.
Customer support team
Extract: Product (name, version), ErrorCode (code, description), Issue (summary, severity), Resolution (steps, outcome).
Create RESOLVES relationships between Resolution and Issue.
Create AFFECTS relationships between Issue and Product.
Skip: boilerplate email signatures, out-of-office replies, automated notification emails.
When the same error code appears in multiple documents, merge into a single ErrorCode entity rather than creating duplicates.
Changing ingestionInstructions after initial ingestion only affects content added after the update. Existing documents in the graph were extracted with the previous instructions (or Along’s defaults). To apply updated instructions to existing content, you need to re-ingest those documents.
Start broad and refine as you observe what the graph captures. Add a few documents, inspect the resulting entities and relationships, then tighten or expand your instructions based on what you see. Iterating on instructions is faster than trying to write perfect instructions up front.