Configuration Reference

Two configuration documents — when to read which. This page is the authoritative complete reference, grouped by the Python class that reads each variable — use it as a lookup table when wiring up your .env. For a purpose-grouped tour (LLM / RAG / DB / observability) read top-to-bottom, see the Configuration guide. When the two pages disagree on a default, this page wins (it mirrors the source code directly).

A full reference of environment variables read by Vedana, grouped by the Python class that reads them.

All three classes below use env_prefix="" and read the same .env file. That means a single env variable like MODEL is read by every class that declares it. The “default” column is the value used only when the variable is absent from .env and the runtime environment — in any real deployment .env.example sets it explicitly.

Vedana Core (`VedanaCoreSettings`)

File: libs/vedana-core/src/vedana_core/settings.py.

Variable	Type	Default	Description
`GRIST_SERVER_URL`	str	—	Grist base URL (`http://grist:8484` or `https://api.getgrist.com`).
`GRIST_API_KEY`	str	—	Grist API key.
`GRIST_DATA_MODEL_DOC_ID`	str	—	DocId of the document with the data model.
`GRIST_DATA_DOC_ID`	str	—	DocId of the document with the data.
`DEBUG`	bool	`false`	Verbose logs and dev features.
`MODEL`	str	`gpt-4.1`	Main LLM model used by `RagPipeline` / `RagAgent` for answer generation and Cypher. This is the value that wins in the standard pipeline — Vedana Core overrides the inner LLM provider on every call.
`ENABLE_DM_FILTERING`	bool	`true`	Enable the data model filtering step.
`FILTER_MODEL`	str	`gpt-4.1-mini`	Model used for filtering.
`JUDGE_MODEL`	str	`gpt-4.1-mini`	Model used for evaluation (LLM-as-judge).
`EMBEDDINGS_MODEL`	str	`text-embedding-3-large`	Embeddings model.
`EMBEDDINGS_DIM`	int	`1024`	Embedding dimensions.
`PIPELINE_HISTORY_LENGTH`	int	`20`	How many recent `comm.*` messages go into the context.
`MEMGRAPH_URI`	str	—	Memgraph Bolt URI.
`MEMGRAPH_USER`	str	—	Memgraph user.
`MEMGRAPH_PWD`	str	—	Memgraph password.

LLM Provider (`LLMSettings`)

File: libs/jims-core/src/jims_core/llms/llm_provider.py.

LLMSettings is the inner provider used by jims-core. It reads the same MODEL env var as VedanaCoreSettings, but if you instantiate jims-core standalone (without vedana-core), no override happens and the fallback default (gpt-4.1-nano) applies.

Which MODEL actually runs? In the standard Vedana stack the two classes read the same env var, so whatever you put in .env (e.g. MODEL=gpt-4.1-mini) is used by both. The defaults differ only as a safety net for standalone jims-core users who don’t set the env var: vedana-core falls back to gpt-4.1 (heavier, better tool-calling), jims-core falls back to gpt-4.1-nano (cheap, OK for non-RAG flows). For all production deployments — set MODEL explicitly in .env.

Variable	Type	Default	Description
`MODEL`	str	`gpt-4.1-nano`	Fallback model for `jims-core` standalone use. Overridden by vedana-core in the standard pipeline.
`EMBEDDINGS_MODEL`	str	`text-embedding-3-large`	Embeddings model.
`EMBEDDINGS_DIM`	int	`1024`	Dimensionality.
`EMBEDDINGS_MAX_BATCH_SIZE`	int	`2048`	Max texts per embeddings batch.
`EMBEDDINGS_MAX_TOKENS_PER_BATCH`	int	`200000`	Max tokens per batch.
`MODEL_API_KEY`	str	`None`	If set, overrides the key for the main model.
`EMBEDDINGS_MODEL_API_KEY`	str	`None`	Same for embeddings.
`OPENROUTER_API_BASE_URL`	str	`https://openrouter.ai/api/v1`	OpenRouter endpoint.

Provider keys (OPENAI_API_KEY, OPENROUTER_API_KEY, GOOGLE_APPLICATION_CREDENTIALS, etc.) are read by LiteLLM from standard env vars.

ETL (`vedana_etl.settings.Settings`)

File: libs/vedana-etl/src/vedana_etl/settings.py.

Variable	Type	Default	Description
`DB_CONN_URI`	str	—	Postgres connection URI for Datapipe.
`GRIST_TEST_SET_DOC_ID`	str	empty	DocId of the document with the golden dataset.
`GDS_TABLE_NAME`	str	`Gds`	Table name in the test set doc.
`TESTS_TABLE_NAME`	str	`Tests`	Test table name in the test set doc.
`TEST_ENVIRONMENT`	str	empty	Environment label for evaluation results.

Database

Variable	Type	Default	Description
`JIMS_DB_CONN_URI`	str	—	Postgres URI for JIMS.
`JIMS_DB_USE_NULL_POOL`	bool	`false`	Disable the connection pool (for serverless).
`JIMS_DB_POOL_SIZE`	int	SQLA default	Pool size.
`JIMS_DB_POOL_MAX_OVERFLOW`	int	SQLA default	Extra connections above the pool.
`CREATE_PGVECTOR_EXTENSION`	bool	`true`	Create the pgvector extension via migration.

Backoffice

Variable	Type	Default	Description
`VEDANA_BACKOFFICE_DEBUG`	bool	`true` (in .env.example)	Dev-only features in the backoffice.

Variable	Type	Description
`TELEGRAM_BOT_TOKEN`	str	Bot token from BotFather.

Sentry

Variable	Type	Description
`SENTRY_DSN`	str	Sentry endpoint. Empty means off.
`SENTRY_ENVIRONMENT`	str	Environment name shown in Sentry.

Datapipe

Variable	Type	Default	Description
`DATAPIPE_PIPELINE`	str	`vedana_etl.app`	Datapipe pipeline entry point.

Prefixes and read order

vedana_core.settings uses env_prefix="" (no prefix) and reads apps/vedana/.env (working directory must be apps/vedana).
vedana_etl.settings — same.
jims_core.llms.llm_provider.LLMSettings — same.
JIMS CLI wrappers use auto_envvar_prefix="JIMS" (click), so the --api-key flag can be provided as JIMS_API_KEY=....

.env.example vs .env.ci-cd

apps/vedana/ ships with two ready profiles:

.env.example — for local development; references db, memgraph, grist (the docker-compose service names).
.env.ci-cd — for CI; uses localhost addresses and test keys.

For production, create your own profile and never commit secret keys (.env is in .gitignore).