Examples

Each example is a complete, commented obsei.yaml in the examples/ folder. The header of each file lists the extras to install, the environment variables it reads and how to run it. Every example is loaded with the real config loader in CI, so the keys match the current release.

git clone https://github.com/obsei/obsei && cd obsei
export OBSEI_DB_KEY="$(openssl rand -hex 24)" OBSEI_PSEUDONYM_SALT="$(openssl rand -hex 24)"
obsei try -c examples/social-listening.yaml        # preview redacted records; nothing stored or sent
obsei run -c examples/social-listening.yaml

Copy a file to obsei.yaml in your project and replace the placeholder ids, hosts and names. Secrets are always read from environment variables.

App reviews in six countries

App Store reviews from the US, UK, India, Brazil, Japan and Germany plus Google Play, classified by a local Ollama model, with negative reviews posted to Slack and multilingual themes.

Sources

• App Store ×6 countries
• Google Play

Your infrastructure

1. Redact: PII and names
2. Classify: decision model or LLM
3. Theme: across languages
4. Store: encrypted

Sinks

• Slack

Extras: google,embeddings

app-reviews-multi-country.yaml

Support tickets with a model cascade

Zendesk, Freshdesk and Intercom (EU region) tickets labelled by a small model; low-confidence results go to a larger in-house model. Bugs become Jira issues, feature requests Linear issues, and everything is exported to Parquet.

Sources

• Zendesk
• Freshdesk
• Intercom

Your infrastructure

1. Redact: PII and names
2. Classify: decision model or LLM
3. Store: encrypted

Sinks

• Jira
• Linear
• Parquet

support-tickets.yaml

Multilingual ticket routing with a decision model

Zendesk and Freshdesk tickets in any language, labelled on a CPU by Julia-1, a 144M decision model served by llama.cpp: spam is filtered out first, then one request per ticket answers sentiment, intent (with descriptions), an urgency score and a yes/no refund question. Answers below their cutoff go to a local Ollama chat model. Ordered routes then pick one destination per ticket: uncertain tickets go to a review channel, confident urgent bugs open a Jira issue and page on-call, billing and refund requests go to the billing channel, and the rest lands in Parquet. A CSV of eight sample tickets in seven languages lets you try it locally: obsei try -c examples/decision-routing.yaml -p sample --enrich --limit 8.

Sources

• Zendesk
• Freshdesk

Your infrastructure

1. Redact: PII and names
2. Classify: decision model or LLM
3. Store: encrypted

Sinks

• Jira
• Slack ×3
• Parquet

decision-routing.yaml

Social listening

Bluesky (English and Portuguese), Hacker News, YouTube comments, a forum RSS feed and a Mastodon hashtag through the REST source, delivered to an internal service as HMAC-signed webhooks.

Sources

• Bluesky
• Hacker News
• YouTube
• RSS
• Mastodon (REST)

Your infrastructure

1. Redact: PII and names
2. Classify: decision model or LLM
3. Store: encrypted

Sinks

• Signed webhook

social-listening.yaml

Trustpilot reviews over REST

Trustpilot Business reviews through the generic REST source, with the API key in a secret header, star ratings mapped and 1-2 star reviews posted to Slack. Field names follow the public business-unit reviews endpoint; verify them against your account’s API version.

Sources

• Trustpilot (REST)

Your infrastructure

1. Redact: PII and names
2. Classify: decision model or LLM
3. Store: encrypted

Sinks

• Slack

trustpilot-rest.yaml

CRM and ITSM over REST

HubSpot tickets (cursor paging on paging.next.after) and ServiceNow incidents (Table API with sysparm_* parameters), classified with custom fields; bugs and outages go to Jira Data Center.

Sources

• HubSpot (REST)
• ServiceNow (REST)

Your infrastructure

1. Redact: PII and names
2. Classify: decision model or LLM
3. Store: encrypted

Sinks

• Jira
• Parquet

crm-and-itsm-rest.yaml

Warehouse round trip

NPS responses read incrementally from Postgres, Snowflake or BigQuery, classified, and written back to a warehouse table without author pseudonyms.

Sources

• SQL warehouse

Your infrastructure

1. Redact: PII and names
2. Classify: decision model or LLM
3. Store: encrypted

Sinks

• SQL warehouse

Extras: sql plus your database driver

warehouse.yaml

Mailbox and file drops

A support mailbox over IMAP and a drop folder for exported CSV and JSON Lines survey and NPS files, with churn risks posted to Slack and a Parquet export.

Sources

• IMAP mailbox
• File drop

Your infrastructure

1. Redact: PII and names
2. Classify: decision model or LLM
3. Store: encrypted

Sinks

• Slack
• Parquet

email-and-files.yaml

Enterprise, air-gapped

vLLM with an Azure OpenAI fallback behind a private egress allowlist, person-name redaction from a local model, regional redaction, local multilingual embeddings, k_anonymity: 10, obsei serve behind an SSO proxy with group-to-role mapping, API tokens, allowed_hosts and a retention job.

Sources

• GitHub Enterprise
• Webhooks
• SQL

Your infrastructure

1. Redact: PII and names
2. Classify: decision model or LLM
3. Theme: across languages
4. Store: encrypted

Sinks

• Jira

· extras mcp,names,embeddings,sql

enterprise-air-gapped.yaml

India, Brazil and Japan

One pipeline per market with market-specific intents, India, Latin America and Asia-Pacific redaction plus name redaction, multilingual themes labelled in one language, and a Slack channel per team.

Sources

• App Store
• Freshdesk
• Intercom
• Bluesky

Your infrastructure

1. Redact: PII and names
2. Classify: decision model or LLM
3. Theme: across languages
4. Store: encrypted

Sinks

• Slack

· extras names,embeddings

regional-india-brazil-japan.yaml

GitHub Actions

A daily workflow: App Store and Google Play reviews, classified, turned into GitHub issues and Slack messages, with the encrypted database kept in the Actions cache. See GitHub Action.

Sources

• App Store
• Google Play

Your infrastructure

1. Redact: PII and names
2. Classify: decision model or LLM
3. Store: encrypted

Sinks

• GitHub issues
• Slack

examples/github-actions

With plugins

These use plugins that are installed separately and allowed with plugins: in the file.

Surveys from Typeform

Typeform survey responses plus a CSV NPS export, classified and grouped into themes, exported to Parquet.

Sources

• Typeform
• NPS CSV

Your infrastructure

1. Redact: PII and names
2. Classify: decision model or LLM
3. Theme: across languages
4. Store: encrypted

Sinks

• Parquet

surveys-typeform.yaml

Alerts in Microsoft Teams

Negative App Store and Google Play reviews posted to a Teams channel.

Sources

• App Store
• Google Play

Your infrastructure

1. Redact: PII and names
2. Classify: decision model or LLM
3. Store: encrypted

Sinks

• Microsoft Teams

alerts-teams.yaml