Configuration

Every backend setting is an environment variable with a sensible default. The backend loads them from backend/app/.env (which overrides the defaults defined in app/config.py). The frontend reads NEXT_PUBLIC_* values from frontend/.env.local.

Change values in both places if you also override them in config.py. The .env file wins at runtime, but the config.py default is the fallback.

#Backend — required

Variable	Description
`SUPABASE_URL`	Your Supabase project URL.
`SUPABASE_SECRET_KEY`	Supabase service role key. The backend uses it server-side to enforce its own access logic (Row Level Security is intentionally bypassed). Keep it secret.
`SUPABASE_JWKS_URL`	JWKS endpoint used to validate access tokens (`…/auth/v1/.well-known/jwks.json`).
`OPENAI_API_KEY`	OpenAI key for embeddings, generation, vision, and the eval judge.
`PINECONE_API_KEY`	Pinecone API key.
`PINECONE_INDEX`	Name of the Pinecone index (must be 1024-dim, cosine).

#Backend — networking

Variable	Default	Description
`ALLOWED_ORIGINS`	`http://localhost:3000`	Comma-separated allowed CORS origins. Set to your frontend origin(s) in production.
`REDIS_URL`	`redis://localhost:6379`	Redis connection for the Arq worker. Required for ingestion and async eval scoring.

#Backend — retrieval & chunking

Variable	Default	Description
`EMBEDDING_MODEL`	`text-embedding-3-small`	Embedding model. Must match the index dimension.
`EMBEDDING_DIMENSIONS`	`1024`	Embedding dimensionality. Must equal the Pinecone index dimension.
`CHUNK_SIZE`	`1800`	Target characters per text chunk (recursive splitter).
`CHUNK_OVERLAP`	`250`	Character overlap between adjacent chunks.
`TABLE_MAX_CHARS`	`3000`	Max size a structured unit (table/slide/row batch) may reach before it is split into prose. Kept above `CHUNK_SIZE` so tables/slides stay whole.
`TOP_K`	`5`	Number of chunks retrieved and fed to the model.
`NEIGHBOR_WINDOW`	`1`	Adjacent chunks stitched around each retrieved chunk (per side). `0` disables.

#Backend — reranking (opt-in)

Variable	Default	Description
`RERANK_ENABLED`	`false`	Set `true` to enable Cohere cross-encoder reranking on `similarity`/`threshold` queries.
`RERANK_MODEL`	`rerank-english-v3.0`	Cohere rerank model.
`RERANK_FETCH_MULTIPLIER`	`4`	Over-fetch factor before reranking (`top_k × this`).
`COHERE_API_KEY`	—	Required when `RERANK_ENABLED=true`.

#Backend — vision / multimodal

Variable	Default	Description
`VISION_MODEL`	`gpt-4o`	Model used for image description and scanned-PDF OCR.
`MIN_IMAGE_BYTES`	`5000`	Embedded PDF images smaller than this are skipped (icons, bullets, rules) to avoid noise.

#Frontend

Variable	Description
`NEXT_PUBLIC_API_URL`	Backend origin the browser calls (e.g. `http://localhost:8000` or your Railway URL).
`NEXT_PUBLIC_SUPABASE_URL`	Supabase project URL (client-side).
`NEXT_PUBLIC_SUPABASE_ANON_KEY`	Supabase anon key (safe for the browser; not the service role key).

The API client normalizes NEXT_PUBLIC_API_URL — if it lacks a scheme it assumes https://. Set the full origin to be safe.

#External resources you must provision

Resource	Requirement
Pinecone index	Serverless, dimension 1024, metric cosine. Name → `PINECONE_INDEX`.
Supabase Storage: `documents`	Private bucket. Holds original uploads; served via 1-hour signed URLs.
Supabase Storage: `images`	Public bucket. Holds extracted PDF figures; their public URLs go into vector metadata for inline source rendering.
Supabase Auth	Email/password enabled.
Supabase tables	`collections`, `collection_members`, `collection_shares`, `collection_documents`, `document_chunks`, `chat_sessions`, `chat_messages`, `ingest_jobs`, `rag_evals`, `query_logs`, `ingest_logs`.
Redis	Reachable at `REDIS_URL`; backs the Arq worker.

#Cost-relevant knobs

Most spend is OpenAI usage. The levers:

TOP_K and CHUNK_SIZE drive prompt size → generation cost per query.
Vision (multimodal PDFs / scanned OCR) is the priciest ingestion path; MIN_IMAGE_BYTES filters trivial images. Prefer native PDFs.
Eval scoring uses the cheaper gpt-4o-mini (two short calls per query).
Reranking adds a small Cohere cost per query when enabled — usually worth it for precision.

See Operations & scaling for the runtime/process view.