title: CloudSlash - Graph Traversal Reference description: DEVI's AI agent doesn't query the graph with SQL. It traverses it generatively: building new paths as it reasons about topology. Translating unstructured hum...

Graph Traversal¶

DEVI's AI agent doesn't query the graph with SQL. It traverses it generatively: building new paths as it reasons about topology.

Context Window¶

Constraint

Standard inference limits (e.g., 128,000 tokens) aren't enough to ingest a .tfstate dump from a 10,000-node architecture. Injecting the raw JSON of the Devi graph into the prompt causes extreme context fragmentation, degrading the model's topological reasoning.

We solve this with a constrained Retrieval-Augmented Generation (RAG) loop that runs strictly against the local BadgerDB index.

Schema-Aware Function Calling¶

Instead of injecting the graph into the prompt, the daemon exposes the API structure as executable functions using the provider's native function-calling constraints (e.g., Gemini's Tools parameter):

// pkg/ai/tools.go
var GraphSearchTool = &llm.Tool{
    Name:        "query_graph",
    Description: "Search the local graph for physical infrastructure matching properties.",
    Parameters: map[string]interface{}{
        "type": "object",
        "properties": map[string]interface{}{
            "node_type": map[string]string{"type": "string"},
            "tag_key":   map[string]string{"type": "string"},
            "tag_value": map[string]string{"type": "string"},
        },
    },
}

When you ask: "Find me all idle databases missing the Owner tag", execution runs in three steps:

1. Inference 1: The LLM parses the intent, recognizes it lacks topological data, and returns a structured {"name": "query_graph", "arguments": {...}} response.
2. Local Execution: The daemon runs the query against the in-memory graph in \(O(1)\) time, returning the specific array of CRNs and telemetry metadata.
3. Inference 2: The daemon appends the results to the context window and re-prompts for natural language synthesis.

Reversibility¶

and Safe Execution

If you request a mutable action ("Quarantine all databases from the previous query"), the LLM generates a proposed cs cleanup command target list.

The daemon never executes LLM output autonomously. The sequence pipes exclusively to the TUI's confirmation buffer. You physically press [Enter] to bind it to a s.a.u saga.

All execution sequences initiated via the LLM are permanently flagged in the s.a.u Audit Ledger with author_identifier: ai-copilot.