API Reference

Provision a ZeroDB project instantly with no signup required. Returns an API key, project ID, and optional claim URL valid for 24 hours.

List all tables in a ZeroDB project. (The old /api/v1/public/zerodb/tables path was deprecated 2026-01-08 and now returns 410 — use this project-scoped path.)

Insert or update vectors in your ZeroDB project. Each vector requires an ID, embedding array, and optional metadata.

Perform semantic similarity search across stored vectors. Returns ranked results with scores and metadata.

Generate text embeddings using AINative's embedding service (16ms average inference). Supports batch up to 512 texts.

Search stored documents by natural language query — embeds your query automatically and returns ranked results.

Store a memory for an agent or user. Supports temporal decay scoring, graph relationships, and vector-backed recall.

Recall relevant memories using hybrid vector + keyword search with temporal decay scoring. Returns ranked results.

OpenAI-compatible chat completions endpoint. Supports Claude, GPT-4o, Kimi K3, GLM-5, DeepSeek V4, and MiniMax M2.7. Works with instant-db API keys.

Provision a new agent swarm project with a named objective. Swarms coordinate multiple AI agents and persist results to ZeroDB.

Register a webhook endpoint to receive real-time event notifications. Supports agent runs, memory updates, and billing events. Signed with HMAC-SHA256.

Write a credit cost node for a task. Set parent_task_id to link it into a multi-agent delegation chain. Costs roll up to the root task for full attribution.

Return the full cost tree rooted at task_id. Uses a recursive CTE to walk parent_task_id links and return every node with its depth. Includes total_credits across all nodes.

Lightweight rollup — returns total credits and node count for a task and all its delegated children. Faster than /tree when you only need the sum.

Combined confidence scoring + hallucination detection. A response passes when confidence >= threshold (default 0.7) AND hallucination_risk < 0.5. Use retry_recommended to decide whether to regenerate.

Ask the LLM to rate 0.0–1.0 how confident it is in the response given the context. Lighter-weight than /evaluate — no hallucination check, no memories required.

Check whether a specific claim is grounded in provided ZeroDB memories. Returns grounded=false with confidence=1.0 when no memories are supplied.