obsei/obsei

obsei – Private, AI‑native feedback analytics

What it is – obsei is an open‑source, self‑hosted platform that ingests customer‑facing text (reviews, tickets, surveys, call transcripts, etc.), redacts personally‑identifiable information, enriches the content with AI models, and makes the results available through webhooks, a web UI (Studio) and an MCP/Claude plugin. All processing happens inside your own infrastructure, so no raw data ever leaves the environment.

Why it matters – Companies often need to understand the voice of the customer but are blocked by privacy regulations (GDPR, CCPA, etc.) or corporate policy that forbids sending raw text to SaaS analytics or public LLM APIs. obsei solves this by:

• encrypting storage and providing built‑in audit logs,
• allowing you to plug in any LLM or lightweight decision model you control (Ollama, OpenAI, Azure, Bedrock, vLLM, etc.),
• automatically redacting PII and optionally pseudonymising authors.

Core components

Component	What it does
Sources	Pull data from CSV/JSONL files, REST endpoints, webhooks, app‑store APIs, GitHub, Hacker News, Bluesky, YouTube, RSS/Atom, SQL warehouses, IMAP mailboxes, any MCP server, Zendesk, Freshdesk, Intercom, Gong, and more (Reddit via community plugin).
Enrichers	Run AI models to classify sentiment, intent, language, custom tags, or any binary/score field. Supports fast CPU‑only decision models (Julia‑1, Clef) or full LLMs via LiteLLM‑compatible endpoints. Low‑confidence results can be escalated to a stronger model or human review.
Sinks	Deliver enriched records to Slack, GitHub issues, Jira, Linear, Parquet files, SQL databases/warehouses, or arbitrary webhooks (HMAC‑signed). Routing rules can direct records based on labels, confidence scores, etc.
Analysis	Offline or multilingual theme clustering (50+ languages) using embeddings, near‑duplicate detection, k‑anonymous views, and a query interface (obsei ask) that returns cited answers. A read‑only Studio UI visualises a knowledge‑graph, trends, decisions and routing panels.
Privacy layer	Region‑aware PII redaction (Americas, Europe, UK, APAC, India, Africa), optional name redaction, salted author pseudonyms, encrypted DuckDB storage, and commands to forget, export or audit data.
Server	obsei serve runs scheduled pipelines, signed webhook intake, MCP HTTP API, Studio UI, role‑based access control and optional SSO‑proxy integration.

Typical workflow

1. Initialize a project (obsei init) – creates obsei.yaml and a sample multilingual dataset.
2. Configure your data sources, choose an enrichment model (e.g., Ollama locally or OpenAI via LiteLLM), and set sinks.
3. Run the pipeline (obsei run) – data are fetched, redacted, enriched, and stored.
4. Serve the system (obsei serve) – pipelines can be scheduled, webhooks accepted, and the Studio UI accessed.
5. Query insights (obsei ask or via the Studio UI) – agents answer questions with citations while respecting privacy filters.

Installation – The project is distributed on PyPI and as a Docker image.

# Python package (install the 1.x pre‑release)
uv tool install "obsei[mcp]>=1.0.0rc1"   # or pip install "obsei[mcp]>=1.0.0rc1"

# Docker (run from your repo directory)
docker run --rm -v "$PWD:/data" -w /data --user "$(id -u):$(id -g)" \
  -e OBSEI_DB_KEY -e OBSEI_PSEUDONYM_SALT ghcr.io/obsei/obsei:1.0.0-rc.3 run

Environment variables OBSEI_DB_KEY and OBSEI_PSEUDONYM_SALT secure the encrypted DuckDB store and author pseudonyms.

Who should use it

• Product teams that need to analyse user reviews across many app stores while staying compliant with data‑privacy laws.
• Support or help‑desk operations that want automated ticket triage, sentiment routing, and escalation without exposing raw tickets to external services.
• Marketing or CX groups performing social listening on public feeds (Reddit, YouTube, Twitter‑like platforms) and needing multilingual theme clustering.
• Enterprises with air‑gapped or highly regulated environments that still want to leverage LLMs for classification, summarisation, or Q&A.

Roadmap snapshot – The 1.0 release candidates (phases 0.1‑0.5) already provide private core, MCP/Claude plugin, enterprise BYO‑model support, theme clustering, and routing. Future work includes live‑connector validation and additional release candidates before the final 1.0.0 launch.

Links – Docs: https://docs.obsei.com/ • Live demo: https://docs.obsei.com/demo/ • Website: https://obsei.com/ • Blog: https://blog.obsei.com/ • Discussions: https://github.com/obsei/obsei/discussions