Skip to main content
This page lists every hard requirement before you start the quickstart. Skipping any of these will cause the deployment to fail at a specific step.

​
Node.js

The backend requires Node.js exactly 22.18.0. The @exulu/backend package includes a preinstall script that checks process.version and exits with a non-zero code if the version does not match:
Use nvm to pin the version. The example repo ships an .nvmrc. The Docker images (Dockerfile.backend, Dockerfile.worker) use node:22.18.0-slim explicitly.

​
Container base image

Both the backend and worker Dockerfiles use node:22.18.0-slim (Debian-based). Alpine Linux is not supported. The ONNX runtime binaries that the knowledge-processing stack depends on are not published for Alpine. Switching to an Alpine base image will cause runtime errors during Python environment setup.

​
Postgres with pgvector

IMP requires Postgres 13 or later with the pgvector extension. The example compose file uses the official pgvector/pgvector:pg17 image, which ships with the extension pre-installed. On managed Postgres (AWS RDS, Azure, Google Cloud SQL), enable the pgvector extension yourself before first boot.

​
A second Postgres database for LiteLLM

The LiteLLM proxy needs its own Postgres database for key management and spend tracking. Set LITELLM_DATABASE_URL to a connection string pointing to this separate database. It can be on the same Postgres server but must be a different database name.
Never set LITELLM_DATABASE_URL to the same database as the IMP application. The proxy runs schema migrations on startup and will attempt to drop tables that belong to the IMP schema.

​
Redis 7

Redis is required when you run workers. Without Redis, BullMQ cannot enqueue or dequeue background jobs (knowledge ingestion, evals, routines). The example compose file uses redis:7. If you deploy without workers (backend only), Redis is optional.

​
S3-compatible storage

File uploads require an S3-compatible bucket. Supported targets include MinIO (local) and AWS S3 (production). The bucket must be created before you start the backend — IMP does not create buckets automatically. Supply the bucket details via the COMPANION_S3_* environment variables.

​
Version pins

The following component versions are tested and pinned. Do not upgrade these without testing — the pins exist because upstream releases have broken compatibility.

​
Notes on the Python pins

  • litellm[proxy]==1.85.1 — the tested version for the embedded proxy. LiteLLM’s API surface changes frequently; upgrade deliberately.
  • prisma==0.15.0 — not included in litellm[proxy] at this version; pinned separately so the LiteLLM database integration works.
  • torch==2.5.1 — pinned because pyannote.audio (used for speaker diarization in transcription) imports torchaudio.AudioMetaData, which was removed in torch 2.6+.
  • transformers>=4.48,<5 — the 5.x line removed a kwarg that pyannote.audio 3.x still passes to hf_hub_download().

​
GPU (optional — Whisper transcription only)

The Whisper transcription server uses WhisperX with CUDA acceleration. Required only if you use the meeting transcription feature.
  • CUDA 12.4
  • NVIDIA driver ≥ 550
  • At least 8 GB VRAM recommended for the default model size

​
Hardware guidance