Public: Harvey legal AI platform

Architectural notes

Harvey runs a multi-provider, auto-routed model stack: an OpenAI partnership
including a custom case-law model (OpenAI Custom Model program) plus o3 and GPT-4.1,
Anthropic Claude Sonnet 3.7, and Google Gemini 2.5 Pro, accessed via Azure OpenAI,
AWS Bedrock, and Google Vertex.

Harvey auto-routes each task to the best model (e.g. one model to research, another
to draft); the routing logic is in-house and not publicly detailed.

RAG uses two stores: LanceDB Enterprise as the primary production vector store (data
kept in customer cloud buckets) and Postgres + pgvector for smaller or public
datasets; embeddings are custom fine-tuned and the model name is not disclosed.

Evaluation uses Harvey-built, open-sourced BigLaw Bench (Answer Score / Source Score)
and a Legal Agent Benchmark (1,200+ tasks, all-pass grading).

Infrastructure is in-house: Kubernetes, an OpenAI-API-compatible model proxy, a model
health tracker, Redis-backed rate limiting, and telemetry exported to Snowflake.

Not publicly documented: any named open-source agent-orchestration framework, the
embedding model, or a Harvey-built guardrail system.