The Anything Engine
14-Class Intent Dispatch

The Anything Engine is Orbiter's intent classifier and dispatch system. A user types any natural language query — "find me investors for our Series A" or "who at LSI should Henry meet?" — and the engine classifies it into one of 14 outcome types, then routes it to a specialized backend tool that runs the appropriate graph query, embedding search, and synthesis pipeline.

The classifier runs on Groq Llama 3.3 70B at near-instant latency. Every response includes a confidence score and reasoning trace, so the router knows when to confirm with the user before dispatching.

FalkorDB is interim. AlloyDB + ScaNN swap pending — same Cypher pattern, higher throughput, 6 vector dimensions per investor.

Every query collapses into one of these canonical classes. The classifier returns the class name, confidence (0–1), and reasoning. LIVE = tool implemented and deployed. PENDING = classified correctly, backend not yet wired.

Classify — Request

POST /api:UgP1h6uR/classify
Content-Type: application/json

{
  "query": "We're raising a $3M seed. Who should we talk to at LSI?",
  "user_id": 15
}

Classify — Response

{
  "class": "find_investors",
  "count": 1,
  "confidence": 0.97,
  "reasoning": "The query explicitly mentions fundraising context ('raising a $3M seed') and requests investor introductions scoped to a specific event (LSI). This maps directly to find_investors with high confidence."
}

Full Dispatch — Request

POST /api:UgP1h6uR/dispatch
Content-Type: application/json

{
  "query": "Find seed-stage VCs who invest in AI infrastructure. $3M round.",
  "user_id": 15,
  "context": {
    "pitch_deck_text": "We build...",
    "live_event_id": 3,
    "zep_thread_id": "thread_abc123"
  }
}

Full Dispatch — Response Shape

{
  "class": "find_investors",
  "confidence": 0.94,
  "results": [
    {
      "master_person_id": 1847,
      "name": "Kai Nguyen",
      "firm": "Gradient Ventures",
      "fit_score": 0.91,
      "rationale": "Gradient led two AI-infra seed rounds in 2024...",
      "why": "Thesis match on AI tooling + infra. Check size $1–5M. 3 portfolio companies in adjacent space.",
      "draft_outreach": "Subject: AI infra seed — Orbiter intro via [name]..."
    }
  ],
  "mem_used": true,
  "process_id": "proc_7c2f4d"
}

// Multi-intent example
{
  "class": "find_investors",
  "count": 2,
  "confidence": 0.88,
  "reasoning": "Query has two intents: find investors (primary) and research the lead investor from their last round (secondary). Dispatching find_investors first, queuing research_person."
}

Prompt Engineering

The classifier prompt is stored in an editable Mintlify doc — never buried in TypeScript. Key elements:

• System: "You are a precise intent classifier for a relationship intelligence platform. Output only valid JSON."
• 14 class definitions with 2–3 canonical examples each
• Explicit handling for ambiguous, compound, and meta-queries
• Instruction to use count>1 rather than hallucinating a single class for multi-intent queries
• Strip markdown fences from response before json_decode (Groq wraps in ```json at <1% rate)

Data Flow

Every dispatch call checks Zep for prior user context. On turn 2 with a vague follow-up query ("show me more like the last one"), the memory layer provides the missing context needed for the classifier to produce a confident routing decision.