Organization knowledge gives Caddie context about your team, workflows, connectors, conventions, and prior outcomes. It helps Caddie make better plans and debug faster while staying inside the organization's access boundary.

Knowledge Sources

Source typeExamples
WorkflowExisting workflow definitions, titles, descriptions, props, and patterns
ConnectorProvider names and available connector types without revealing secrets
MemoryUseful facts extracted from prior Caddie interactions or workflow outcomes
ManualCurated notes added by the organization
Run patternRepeated runtime behavior, failures, or successful execution patterns

Retrieval Flow

graph LR
Context[User request] --> Query[Build retrieval query]
Query --> Search[Search org knowledge]
Search --> Rank[Rank relevant chunks]
Rank --> Agent[Caddie context]
Agent --> Plan[Workflow plan or answer]

What to Store

Good organization knowledge is operational and non-secret:

StoreAvoid
Naming conventionsAPI tokens
Preferred chains and assetsWallet signing material
Approved Slack channelsWebhook secrets
Wallet policy notesRaw connector credentials
Customer or protocol routing metadataSensitive personal data not needed for automation
Incident and runbook summariesInternal secrets or privileged prompts

Store facts like "large transfers require review" or "alerts go to #protocol-ops." Keep secret material in connectors, API keys, service accounts, or wallet controls.

Caddie Profile

An organization Caddie profile can describe preferences such as tone, default chains, operational rules, naming conventions, review requirements, and workflow style. Keep it concise and directive.

md
Default chain: BasePreferred alert channel: protocol-opsBefore wallet transactions: add a policy branch and require a test runNaming: use verb-object names for workflow titles

Maintenance

  1. Remove stale facts when provider accounts, wallet policy, or org structure changes.
  2. Prefer short, specific knowledge entries over long internal documents.
  3. Avoid copying secrets or private incident data into knowledge.
  4. Use successful templates as reusable knowledge instead of repeatedly explaining the same pattern.
  5. Review knowledge when Caddie starts making outdated assumptions.
Ask a question... ⌘I