Decentralized AI verification network that breaks LLM outputs into atomic claims and runs multi-model consensus across verifier nodes. SDK exposes verified inference and flow orchestration.
- 01AI output verification
- 02multi-model consensus
- 03hallucination filtering
- 04verified LLM agents
- 05RAG with fact-checking
- pip install mira-sdk
- pnpm add mira-network
| Variable | Scope | Description |
|---|---|---|
| MIRA_API_KEY | Server | Mira Network API token created in the Mira Console. |
Use Mira Network to wrap LLM calls with consensus verification. From Python: `from mira_sdk import MiraClient; client = MiraClient(api_key=os.environ['MIRA_API_KEY'])`. Call `client.generate(model='mira/llama3.1', messages=[...])` for verified inference, or build a Compound Flow with `client.flows.create(...)` to chain elemental flows. For verification, call the Mira Verify API on any LLM output: each claim is voted on by multiple models and a supermajority returns `verified=true`. Treat unverified claims as untrusted and surface them to users.
- ⚑Verification adds latency — each claim is fanned out to multiple verifier nodes, so expect 2-10x slower responses than a raw LLM call; stream where possible.
- ⚑Consensus is probabilistic, not cryptographic — a supermajority of biased models can still ratify a wrong claim; do not treat `verified=true` as a hard guarantee for safety-critical flows.
- ⚑Network reliability depends on verifier node availability; design fallbacks for partial-consensus or timeout responses.
- ⚑API credits deplete per token verified — long contexts can blow through budgets quickly; sample claims rather than verifying every token.
- ⚑Models available on Mira (GPT-4o mini, Llama 3.x, DeepSeek-R1) are a fixed roster — you cannot bring arbitrary fine-tunes through the consensus pipeline.