feat: preconfigured Ollama provider with granite4:350m

[Empreiteiro]

Summary

Turnkey Ollama experience: starting Langflow auto-configures the Ollama provider in Model Providers with granite4:350m ready to use. Works for both make run_cli (using docker or podman, with optional Homebrew autoinstall fallback) and a new dedicated Compose stack.

granite4:350m is the bundled default — ~676 MB on disk, supports tool calling, much quicker to bootstrap on a fresh machine than the 2 B variants. The model name is configurable via LANGFLOW_OLLAMA_MODEL.

Wiring (so the provider shows up in the UI)

• Add OLLAMA_BASE_URL to VARIABLES_TO_GET_FROM_ENVIRONMENT so it gets promoted to a per-user Variable on first login — this is what marks the provider as enabled in the Model Providers UI.
• Register granite4:350m in the Ollama catalog (tool_calling=True) so it appears in the dropdown.
• run_cli and run_clic export OLLAMA_BASE_URL=http://localhost:11434 for the Langflow process.

Bootstrap script (scripts/setup/ollama_bootstrap.sh)

Idempotent. Resolution order:

1. Something already responds on :11434 → use it.
2. A container engine is available — docker OR podman, autodetected — start (or restart) langflow-ollama and <engine> exec an ollama pull granite4:350m. The Ollama image is pinned to docker.io/ollama/ollama:latest so it resolves under Podman's default registry policy. Engine can be forced with LANGFLOW_OLLAMA_ENGINE=docker|podman.
3. Local ollama CLI installed → ollama serve in the background, then ollama pull.
4. macOS + Homebrew → brew install ollama, then go to step 3. Skippable with LANGFLOW_OLLAMA_NO_INSTALL=1.
5. None of the above → fail loudly with concrete install instructions (Podman / Docker / Ollama) instead of silently letting Langflow start with a dead OLLAMA_BASE_URL. Escape hatch: LANGFLOW_OLLAMA_OPTIONAL=1 make run_cli (warns and continues).

Make targets

• make ollama_up — runs the bootstrap. Listed as a dependency of run_cli / run_clic.
• make ollama_down — stops/removes the langflow-ollama container (engine-agnostic — works with podman too). Keeps the model volume.

Compose stack — docker_example/ollama-granite/

Standalone, runs end-to-end with one command:

docker compose up      # or: podman compose up

Services: langflow + postgres + ollama + ollama-init (one-shot sidecar that pulls granite4:350m then exits). NVIDIA GPU block included but commented.

How a user runs it

Local (Podman / Docker / no engine all handled):

make run_cli

Compose:

cd docker_example/ollama-granite && docker compose up

Open http://localhost:7860. First start downloads granite4:350m (~676 MB). Ollama appears in Model Providers already enabled.

Verified on a dev machine (macOS + Podman)

• podman ps → container langflow-ollama Up
• GET http://127.0.0.1:11434/api/tags lists granite4:350m (676 MB)
• POST /api/generate to granite4:350m returns a real completion
• After first login on a fresh DB, the Ollama card in Model Providers is enabled and the model is selectable
• make run_cli exports OLLAMA_BASE_URL and Langflow log confirms Set default_fields for non-secret variable OLLAMA_BASE_URL

Test plan

• make run_cli on a Podman host → container created, model pulled, Langflow starts wired up
• make run_cli on a Docker-only host → same path, engine auto-detected as docker
• make run_cli with ollama serve already on :11434 → bootstrap detects it and skips the engine path
• make run_cli on macOS without Docker/Podman/Ollama → autoinstalls Ollama via Homebrew, then starts ollama serve
• make run_cli on Linux without Docker/Podman/Ollama → fails loudly with install instructions (no silent pass-through)
• LANGFLOW_OLLAMA_OPTIONAL=1 make run_cli on a host without anything → warns and continues
• After first login on a fresh DB, the Ollama card in Model Providers is enabled and granite4:350m is available
• docker compose up and podman compose up in docker_example/ollama-granite/ both bring the stack up end-to-end
• make ollama_down cleans up the container under both docker and podman (without wiping the model volume)
• Backend variable-service seed tests still pass (test_initialize_user_variables__create_and_update, etc.)

Files

• src/lfx/src/lfx/services/settings/constants.py — adds OLLAMA_BASE_URL to env-seed list
• src/lfx/src/lfx/base/models/ollama_constants.py — registers granite4:350m
• Makefile — ollama_up / ollama_down targets, wires into run_cli / run_clic
• scripts/setup/ollama_bootstrap.sh — engine-agnostic bootstrap
• docker_example/ollama-granite/{docker-compose.yml,README.md} — turnkey stack

Configuration knobs (env vars)

Variable	Default	Purpose
OLLAMA_PORT	11434	Port to expose Ollama on
OLLAMA_HOST	127.0.0.1	Host the script probes
LANGFLOW_OLLAMA_MODEL	granite4:350m	Model to ensure-pull
LANGFLOW_OLLAMA_ENGINE	(auto)	Force docker or podman
LANGFLOW_OLLAMA_NO_INSTALL	0	Disable Homebrew autoinstall fallback
LANGFLOW_OLLAMA_OPTIONAL	0	Never exit non-zero (warn-and-continue)

🤖 Generated with Claude Code

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: d9f37250-6d20-4aa6-8adf-f1086a42f1f1

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}