Obsei

obsei

Voice of Customer without violating privacy. The self-hosted, open source alternative to Enterpret and Unwrap that strips PII before any LLM sees it.Bring your own sources, models and agents.

Docs · Live demo · Website · Blog · Discussions

---

Teams want to understand what customers say in reviews, tickets, calls and surveys, but often can't send that text to another SaaS or a public LLM. obsei runs entirely inside your own infrastructure:

• Private by design. PII is redacted (and optionally names) and authors are pseudonymised at ingest. Encrypted database, air-gapped by default, no telemetry, erasure and an audit log built in.
• Bring your own. Your sources, your keys and your model: a zero-shot classifier that runs in process by default, a decision model such as Julia-1, or Ollama, vLLM, Azure OpenAI, OpenAI, Bedrock and Vertex through LiteLLM.
• AI-native. A read-only MCP server and a Claude plugin, so agents answer questions about customers with cited, privacy-filtered evidence.
• Every language. Feedback is classified and quoted in its own language. With obsei[embeddings], a local multilingual model groups the same issue across 50+ languages; without it, the offline embedder groups by wording, so each language forms its own themes.

obsei provides controls that support compliance with laws such as the GDPR, India's DPDP Act, Brazil's LGPD and California's CCPA. It makes no compliance claims; you remain the data controller.

[!NOTE] 1.0 is a new codebase, not compatible with 0.0.x, and is in pre-release: install it with the >=1.0.0rc1 specifier. A plain pip install obsei still gives 0.0.15, whose code lives on the legacy/0.0.x branch. See Migrate from 0.0.x.

Quickstart

uv tool install "obsei[mcp]>=1.0.0rc1"   # or: pip install "obsei[mcp]>=1.0.0rc1"
mkdir voc && cd voc
obsei init                             # obsei.yaml plus a 10-language sample dataset
export OBSEI_DB_KEY="$(openssl rand -hex 24)" OBSEI_PSEUDONYM_SALT="$(openssl rand -hex 24)"
obsei try                              # preview redacted records; nothing stored or sent
obsei run                              # fetch, redact, enrich, deliver, store
obsei serve                            # scheduled pipelines, webhooks, MCP and Studio
obsei doctor                           # environment, encryption, egress mode, plugins

Uncomment the classify enricher in obsei.yaml to label sentiment, intent and language with the local classifier: install obsei[local] and run obsei models download once; it runs on the CPU with no model server. A decision model on llama.cpp or any OpenAI-compatible endpoint also works: Ollama, vLLM, Azure OpenAI, OpenAI, Mistral, or a LiteLLM proxy for Bedrock and Vertex (benchmark). Public endpoints are refused until you choose OBSEI_EGRESS_MODE=private (with OBSEI_EGRESS_ALLOW) or hybrid.