LLM Authorization#

Adjust the security level for LLM interactions controlling what kind of data is sent to the LLM during database interactions:

  • NO_LLM: No LLM connected; nothing is sent to an LLM.

  • METADATA: Only questions and database schemas are sent to the LLM (e.g., for query generation).

  • DATA: LLM can see table values.

Configure each setting by specifying LLM_PRIVACY_LEVEL=... in data/custom.env.

To Disable LLM#

LLM_PRIVACY_LEVEL=NO_LLM

Authorization & Data Flow#

LLM Compliance: Follow oversharing and AI-decision-making policies

Supported LLMs: OpenAI or any LangChain-compatible (Azure OpenAI, AWS Bedrock, Llama2, Falcon, Mistral, …) can be connected; typically configured to NOT learn from user queries to avoid data leakage

Privacy Level Settings#

Set LLM_PRIVACY_LEVEL=...

  • NO_LLM: No LLM connected; nothing sent to LLM; no AI decision making

  • METADATA: User query, database schema, and previous outputs go to LLM… but no table values. Minor leakage to LLM such as analyst intent, special values in queries, etc. AI generates queries and code, but analyst checks, and does all data interpretation.

  • DATA: Data table contents sent to LLM for analysis, e.g., Table AI agent

Learning & Context#

Louie learns from usage too

Louie builds a database of context to learn smarter queries over time (RAG)

  • Currently: Query history limited to same thread

  • Upcoming: Control widening scope to organizational sharing unit (RBAC), e.g., personal account vs team