How Vashist's AI works.

Vashist routes requests to AI models via OpenRouter, which lets us select between several providers depending on the task (chat, embeddings, rephrasing) and the context. Primary models come from the Claude family (Anthropic) and the GPT family (OpenAI), with other families used selectively.

Specific model names may rotate as new versions become available. Configuration is managed internally. You can request a log of the models used in your chat history through support.

For PRO users, Vashist uses Mem0 to keep context across conversations — remembering your dosha, your practices, and the preferences you've shared. Each user has their own isolated memory space; it is never crossed between accounts.

Data location: memory vectors live in our PostgreSQL database with pgvector, on servers located in the European Union.

Deletion: there is currently no self-service switch to disable memory, but you can request immediate deletion through support. When your account is downgraded from PRO to BASIC, persistent memory is also removed. A direct opt-out control is on our roadmap.

When Vashist answers you, the answer is grounded in sources — not invented. The system uses RAG (Retrieval-Augmented Generation) over a curated knowledge base, which includes:

When an answer uses a specific source, it is cited with a link to the document. If there is no citation, the answer comes from the model's general knowledge — something that may be wrong and should be verified.

Every AI call is logged and analysed with Langfuse — an observability tool for AI systems. It lets us monitor latency, cost, error rate, and perceived response quality over time.

Logs contain the input, output, model used, and token count. They do not contain additional personally identifiable information — only the content you shared in the conversation. Maximum retention is 90 days.