Evo.ai

Evo.ai is EvoPlatform's Retrieval-Augmented Generation (RAG) layer. Point it at your data, plug in the model of your choice — a cloud frontier model or a local/offline one — and get a tenant-isolated assistant that answers from your data, with citations, and refuses questions your records don't cover.

It is an API-first service: a FastAPI backend with a built-in chat UI at /ui, Qdrant for vectors, and a pluggable LLM/embedding layer. Other Evomedia products call it with a tenant JWT; the evo.ehs connector and its "Ask AI" panel are the first consumers.

What it does

• Answers only from your data. A grounding policy and a relevance gate keep the assistant on-topic: off-topic questions are refused before any LLM call, and every answer cites the source chunks it used.
• Pluggable models, per tenant. Run fully local with Ollama, or point a tenant at any LiteLLM provider (Anthropic, OpenAI, Gemini, Mistral, Groq, Azure, Bedrock, …). Provider API keys are encrypted at rest.
• Hybrid retrieval. Dense vectors plus BM25 sparse vectors with fusion, which materially improves recall on IDs, names, and codes over dense-only.
• Strict multi-tenant isolation. A physical Qdrant collection per (tenant, collection) and a tenant_id metadata filter on every query, with the tenant id taken only from verified JWT claims.
• Source connectors. Sync a Confluence space, a Jira project, or a evo.ehs site into a collection, with incremental sync and deletion handling.
• Structured analytics. For evo.ehs tenants, aggregate questions ("how many permits are open?") are answered by a validated, read-only SELECT over tenant-scoped views — the kind of counting/ranking question that top-k chunk retrieval can't answer.

How the pieces fit

Layer	What it is
API	FastAPI app — config, ingest, query, sources, and admin routers
Chat UI	Served at /ui; talks to /query, renders cited chunks with scores
Vector store	Qdrant — one collection per (tenant, collection), hybrid dense + sparse
Models	Ollama (local) or any LiteLLM provider, per tenant, for both LLM and embeddings
Metadata store	SQLite — tenant model config, collections, doc registry, sources, service keys, tenant links
Auth	JWT verified against EvoPlatform's JWKS; tenant_id from claims

Where to go next

• Getting started — run the dev stack and ask your first question.
• Configuration — settings, models, embeddings, retrieval, and guardrails.
• Ingesting data — files, structured records, document identity, and the watched-folder agent.
• Querying — the query endpoint, citations, multi-turn, and the relevance gate.
• Connectors — Confluence, Jira, and the evo.ehs connector with site auto-discovery and analytics.
• Multi-tenancy & security — isolation, auth, service keys, tenant links, and encryption.
• API reference — every endpoint.
• Operations — deployment and the quality gates.