Configuration

sqldbagent is configured with Pydantic Settings and .env, with nested settings exposed under the SQLDBAGENT_ prefix.

Core Areas

  • Datasources: direct datasources config or convenience Postgres, Postgres demo, MSSQL, and SQLite fields.

  • Agent orchestration: model selection, middleware controls, tool limits, checkpoint persistence, and long-term memory store behavior.

  • Retrieval: Qdrant connection details, collection naming, and search defaults.

  • LangSmith: tracing enablement, project naming, workspace selection, and default tags.

  • Prompt/runtime ergonomics: prompt enhancement toggles, tiktoken-backed token estimation with deterministic fallback, and live explored context persisted under the artifact root.

Datasource Convenience Fields

The standard local variables are:

POSTGRES_HOST
POSTGRES_PORT
POSTGRES_DB
POSTGRES_USER
POSTGRES_PASSWORD

POSTGRES_DEMO_HOST
POSTGRES_DEMO_PORT
POSTGRES_DEMO_DB
POSTGRES_DEMO_USER
POSTGRES_DEMO_PASSWORD

MSSQL_HOST
MSSQL_PORT
MSSQL_DATABASE
MSSQL_USER
MSSQL_PASSWORD

SQLITE_PATH

LangGraph Checkpointing

Use Postgres checkpointing for durable agent threads across runs:

SQLDBAGENT_AGENT_CHECKPOINT_BACKEND=postgres
SQLDBAGENT_AGENT_CHECKPOINT_AUTO_SETUP=true
SQLDBAGENT_AGENT_CHECKPOINT_PIPELINE=false

If an explicit checkpoint URL is not provided, sqldbagent synthesizes one from the standard POSTGRES_* fields.

Query Safety Defaults

  • Datasources default to read-only execution behavior.

  • Writable execution is only available when the datasource safety policy enables it explicitly through allow_writes=true and the caller requests writable access.

  • Postgres uses read-only transactions, SQLite uses PRAGMA query_only, and MSSQL requests ApplicationIntent=ReadOnly while still relying on the shared SQL guard as the hard boundary.

Prompt Artifacts

  • Prompt bundles and prompt enhancements are stored under the artifact root.

  • Prompt exports cache token estimates for the base prompt, final prompt, enhancement text, and prompt delta.

  • Prompt enhancement artifacts also cache per-layer estimates for generated context, live explored context, user context, business rules, direct effective-prompt instructions, and answer-style guidance.

  • Live prompt exploration can save additional read-only discovered context back into the enhancement artifact for the same datasource/schema.

LangSmith Tracing

LangSmith tracing is optional and local-key friendly:

SQLDBAGENT_LANGSMITH__TRACING=true
SQLDBAGENT_LANGSMITH__PROJECT=sqldbagent-dev
SQLDBAGENT_LANGSMITH__TAGS=sqldbagent,local,dashboard
LANGSMITH_API_KEY=...
LANGSMITH_ENDPOINT=https://api.smith.langchain.com
LANGSMITH_WORKSPACE_ID=...

When langgraph.json points at .env, the same tracing config is available to langgraph dev and dashboard turns without duplicating configuration.

Example

See the repo-root .env.example for the supported local shape.