How to describe Gravitre to buyers
Gravitre does not require customers to migrate or centralize business data into a separate data platform. Information stays in your existing systems (CRM, drive, ERP, etc.) and is accessed through authorized connectors.
Gravitre does store operational data needed for AI features — conversation history, knowledge-base indexes, embeddings, workflow records, and audit logs — in your org-scoped database until you delete them or your retention policy applies.
For each AI request, Gravitre sends the selected model provider only what is required for that turn: instructions, your question, recent dialogue, retrieved text passages, connector tool results for that step, and org metadata — not full documents, embedding vectors, or bulk exports.
Avoid this claim
Do not say "Gravitre doesn't store customer data." Security reviewers will ask where chat history and knowledge indexing live. The accurate answer is qualified storage for AI operations, with connector-first access to systems of record.
Is data cached?
Yes. Gravitre caches retrieval results, query embeddings, org context snapshots, and some assistant responses to improve latency.
| Cache type | Typical duration |
|---|---|
| Search / retrieval results | 5 minutes |
| Query embeddings | Up to 7 days |
| Org context snapshot | 60 seconds |
| Tier-0 cached answers | 1 hour |
Caches are performance layers, not the system of record. When Redis is enabled, treat it as ephemeral sensitive storage.
Are prompts stored?
Yes — Chat and assistant threads persist user prompts and assistant replies in your org database until you delete the conversation.
Billing telemetry (model_calls) records token usage and cost — not full prompt or response text.
Are embeddings stored?
Yes — in Supabase Postgres (pgvector) for knowledge search and agent memory. Default embedding model: OpenAI text-embedding-3-small (1536 dimensions).
Embedding vectors stay in your database. They are not sent to chat models at inference time.
Are workflow outputs stored?
Yes — workflow runs store step inputs/outputs (JSON snapshots on the legacy workflow path; lighter metadata on the newer runs API). This supports audit, replay, and debugging.
Are conversation histories stored?
Yes — until deleted via Chat history or the conversations API. Active model context uses recent turns plus an optional rolling summary when threads grow long.
Enterprise admins can configure dialogue retention settings via API; automated chat purge jobs should be confirmed with your contract / deployment.
What exactly is sent to the model?
| Sent to the LLM | Not sent to the LLM |
|---|---|
| System instructions and safety rules | Full source documents |
| Your question (PII-redacted by default) | Embedding vectors |
| Recent chat turns | Bulk CRM/ERP exports |
| Retrieved passage text (chunk excerpts) | OAuth tokens or API keys |
| Source titles / scores / metadata | Entire connector datasets |
| Connector tool results for the current step | Internal infrastructure secrets |
| Org status metadata (integrations health, counts) |
Chunking: Ingest splits documents into ~1,000-character segments. At question time, only the best-matching passages (typically truncated further to 500–1,200 characters each) are included in the prompt.
Can admins disable specific models?
Yes.
- Deployment-wide: disable all AI (
DISABLE_AI). - Per org: model allowlist / blocklist via Settings API (
/api/settings/model-policy) — restrict providers (OpenAI, Anthropic, Gemini) and/or model names. - Rate limits and hard budgets provide additional spend controls.
See Security overview and Enterprise for residency and SIEM options.
Can data be redacted before leaving the environment?
Partially — by default, yes for user and retrieved content.
Before external LLM calls, Gravitre applies best-effort PII redaction (email, phone, SSN patterns, credit card patterns) when AI_PII_REDACTION_ENABLED is on (default). User content is wrapped in untrusted-input fences for prompt-injection hardening.
System instructions are not redacted. Tool observation payloads in agent loops may contain connector fields as returned by the vendor API — review connector scopes and approvals accordingly.