feat: Add mlflow filter pipeline

[WeichenXu123]

MLflow Filter Pipeline

Adds a filter pipeline that integrates https://mlflow.org/ tracing with Open WebUI, enabling observability for multi-turn chat sessions.

What it does

• inlet: Captures the last user message and session context before each request
• outlet: Logs a complete trace per turn — user input, assistant response, model name, and token usage — grouped under the same session in the MLflow UI

Features

• Multi-turn session grouping — all turns of a conversation are linked via mlflow.trace.session, viewable with "Group by session" in the MLflow UI
• Per-turn tracing — each request/response pair is logged as a separate MLflow trace with latency and status
• Token usage tracking — input/output token counts are captured when provided by the backend, automatically aggregated at the trace level
• User attribution — traces are tagged with the authenticated user's email via mlflow.trace.user

Requirements

• MLflow tracking server running and accessible: mlflow server --host 0.0.0.0 --port 5000
• requirements: mlflow>=2.14.0

Configuration (Valves)

• mlflow_tracking_uri (default: http://localhost:5000) — MLflow tracking server URI
• mlflow_experiment_name (default: open-webui) — Experiment name in MLflow
• debug (default: false) — Enable debug logging

Setup

1. Start MLflow: mlflow server --disable-security-middleware
2. Upload this pipeline in Open WebUI → Admin → Pipelines
3. Set MLFLOW_TRACKING_URI env var (Docker users: http://host.docker.internal:5000)
4. Open the MLflow UI and enable "Group by session" to view full conversations

end-to-end test in local machine:

Launch Mlflow server

MLFLOW_LOGGING_LEVEL=DEBUG mlflow server --disable-security-middleware

Launch Openweb UI server

open-webui serve

Launch pipeline service by docker

1. build a dockerfile and save it as Dockerfile.mlflow in current directory.

FROM ghcr.io/open-webui/pipelines:main
RUN pip install --no-cache-dir mlflow

2. build image:

docker build -f Dockerfile.mlflow -t pipelines-mlflow .

3. launch the docker container:

docker run -p 9099:9099 --add-host=host.docker.internal:host-gateway -v pipelines:/app/pipelines --name pipelines --restart always -e MLFLOW_TRACKING_URI=http://host.docker.internal:5000/  -e DEBUG_MODE=true pipelines-mlflow

OpenWebUI pipeline configure:

OpenWebUI chat:

MLflow traces:

single traces:


chat session:

[Copilot]

Pull request overview

Adds a new example filter pipeline that integrates MLflow tracing with Open WebUI to provide per-turn traces grouped by multi-turn chat session.

Changes:

• Introduces an MLflow filter pipeline with configurable valves (tracking URI, experiment name, debug).
• Captures the latest user message on inlet and logs the corresponding assistant response + token usage on outlet.
• Groups turns into a single MLflow session via mlflow.trace.session and tags traces with the authenticated user email.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[WeichenXu123]

CC @tjbck would you take a look ? thanks!