Data & privacy
Where your data lives, what is sent to AI providers, how personal data is protected, and what you can delete.
This page explains what happens to your data when you use Tattva. Share it with your security team before you connect a sensitive source.
Where your data lives
- BigQuery tables stay in your warehouse. Tattva reads their schema and a profile of each column, and runs queries there when it answers. Rows are not copied into Tattva's database.
- Uploaded files are stored by your deployment (in Google Cloud Storage when your operator has configured it). Tattva queries them in place when it answers.
- Documents you upload (
.md,.txt,.html) are stored as text so they can be searched. Metric documents (.mdor.txt) are also read, word for word, by the ontology build. - Conversations: your questions, the answers and every step in between are stored so you can reopen them from the Sessions drawer and audit them in the Trail.
- Column profiles: for text columns with up to 1,000 distinct values, the 20 most common values are stored so answers can match the names people use. This is never done for columns marked as personal data.
- Embeddings (numeric representations used for search) and audit events are stored in Tattva's database.
What is sent to AI providers
When you ask a question, parts of your data are sent to an AI provider to compute the answer. That includes:
- your question and the recent conversation;
- a description of your data: table and column names, their meanings, metric definitions;
- the rows returned by the queries Tattva runs while answering, after the personal-data rules below are applied.
| Provider | What it receives | Why |
|---|---|---|
| Anthropic (Claude) | The items above | The main reasoning models |
| OpenAI | The same, when a task is routed to it | Fallback and some tasks |
| Google (Gemini) | The same, when your operator routes a task to it | Some tasks |
| Voyage AI | Text of document passages, ontology entries and search queries | Search embeddings |
| Google (sign-in) | Your email and basic profile | Signing in |
Which models handle which work is listed on Settings › Integrations. See Integrations.
Anthropic may cache stable parts of a prompt for a few minutes to answer faster, under its prompt caching terms.
How personal data is protected
Each workspace member has a PII clearance tier, set by an Admin or Owner: none (the default), aggregated, unsafe aggregates, or full. Tattva applies it on every question:
- Metrics and columns classed as personal data are left out of what the model sees, unless your tier allows them.
- A query the agent writes that reads a column marked
piiis refused before it runs unless your tier covers it, and the refusal is logged. - For members with aggregated or unsafe-aggregates clearance, BigQuery queries whose groups would describe too few people are rewritten or have those rows suppressed, so no individual can be picked out.
- Answers for members with none or aggregated clearance are scanned for emails, phone numbers and card numbers, which are redacted: in the text, and in the tables the answer's charts, chart images, saved reports and Copy tables are drawn from.
- Grants and revocations are recorded in the audit log.
This only works if sensitive columns are marked. Data analysts review and correct column sensitivity on each source's Schema tab. See Sensitive data.
Data you give to a custom tool is sent to that tool's endpoint. See Custom tools.
Data residency
Tattva runs where your operator deploys it. The AI providers above are US-based by default. Provider regions can't be set per workspace or per project. If you have residency requirements, raise them with your operator before connecting sources.
Encryption
- Calls to AI providers use HTTPS.
- Stored data relies on your cloud provider's managed encryption at rest.
- API keys are stored only as hashes. The full token is shown once, when you create it.
- Sign-in sessions use long random tokens and last 30 days.
- TLS for the site itself is configured by your operator at the load balancer.
Deleting data
- A source: Delete source removes it from the database and from storage, including every version.
- A project: Delete project on Settings › Project removes its conversations, decisions, trackers and uploaded sources.
- A workspace: owners can delete it from Settings › Workspace.
- Signing out ends your session.
Embeddings already computed by Voyage AI are held under Voyage's own retention terms; deleting a source doesn't ask Voyage to delete them.
There is no self-service account deletion. Ask your platform operator.
Sign-in
You sign in with Continue with Google. Tattva doesn't manage passwords or multi-factor settings for production accounts; your Google account's settings apply.
Third-party processors
| Processor | Data | Purpose |
|---|---|---|
| Google Cloud (Storage, Cloud SQL, BigQuery) | Platform data | Hosting |
| Google sign-in | Email, basic profile | Authentication |
| Anthropic | Questions, conversation, data descriptions, query results | Reasoning |
| OpenAI | The same, when routed | Fallback and some tasks |
| Google Gemini | The same, when routed | Some tasks |
| Voyage AI | Text passages | Search embeddings |
For your security review
- Read this page and Sensitive data.
- Sample the data your team will connect. If it holds personal data, check that those columns are marked before anyone asks questions.
- Confirm with your operator which AI providers are enabled; Settings › Integrations shows them.
- Review the audit log to check it records what you need.