Compounding Inference Swarm Engine

[iamcapote]

---

---

Compounding Inference Swarm Engine

Multi-pass reasoning loop with exponential semantic “compounding,” self-reflection, and final synthesis.

Primary use-cases: technical Q&A, quick fact checks, light calculation, abstract problem breakdown.

---

How it Works

1. Chat Trigger – Query Input node exposes a public chat interface.
2. State Init – Initialize Inference State seeds a global static object:
   • unique sessionId
   • prompt, iteration 0, golden-ratio harmonic factor φ
3. Prompt Meta-Generation – Build RAG Prompt → Prompt Generator Agent (LLM + Think tool) crafts a deeper, orthogonal prompt; Clean Agent Output strips any JSON/markdown.
4. Prompt Stacking – Restore State After Prompt pushes the new prompt into history.
5. Core Inference – Prepare Inference Prompt → Core Inference Agent (LLM + Reflection + Memory window 20) performs analysis; Clean Inference Output sanitises; Parse & Compound updates:
   • depth = depth × φ
   • dimensionality ++
   • compoundingFactor = e^{π·depth/dimensionality}
   • confidence asymptotically → 0.95
6. Loop Control – Continue Compounding? repeats steps 3-5 while
   confidence < threshold AND iteration < maxIterations (8).
   Final Synthesis – when the loop exits, Prepare Synthesis builds a global recap; Synthesis Agent produces the answer; Clean Synthesis Output returns raw text to the user.

---

Usage Tips

• Ask any question; the engine self-propagates deeper prompts until convergence.
• Plain-text only; structured output is stripped.

---

Limits & Errors

• Iterations: ≤ 8
• Confidence stop: 0.95
• Token limits: 3 k (synthesis/core), 1.5 k (prompt gen).
• Any tool/LLM error bubbles up as thrown JS error.

---

Permissions: Public

Login for demo

• user: testuser
• pass: testuser1234

---

---

Workflow Details