// AI ORCHESTRATION
Calling the model is easy.
Everything around it is the work.
Getting the right data into the prompt, handling an answer that isn't what you asked for, and routing the result to the next system: that is a pipeline, and pipelines are what Apache Hop runs. The model is a connection setting.
● Any provider
OpenAI, Anthropic, Mistral, Ollama, Hugging Face, or any OpenAI-compatible endpoint
● Local if you want
With Ollama, prompts never leave your network
● Off by default
Nothing goes to a model until you set up a provider
● Open source
Apache Hop is an Apache Software Foundation project
// APACHE HOP
Language models as pipeline steps
Every AI step is an ordinary Hop transform: rows in, rows out, next to the transforms that read your sources and write your targets. One AI Provider connection serves them all.
APACHE HOP
Language model chat
A prompt per row, built from the row's fields, and the answer added back to the row. Classify tickets, summarise articles, draft replies.
APACHE HOP
Structured extract
Named, typed fields out of free text: a support ticket becomes a severity and a product, a contract becomes a renewal date and a value. Columns you can filter, join and aggregate.
APACHE HOP
Retrieval
Text chunker, Embed text, then pgvector upsert to store your documents, and pgvector search to find the passages closest to a question. A regular PostgreSQL connection is all the vector store needs.
APACHE HOP
AI Assistant in Hop GUI
Ask about the pipeline or workflow you have open. The assistant proposes changes; nothing is applied until you review it. Off until you enable it.
// PROVIDERS
Bring the model you have
Switching models means changing the AI Provider connection. The prompt, the parsing and everything downstream stay the same.
| PROVIDER | WHERE IT RUNS | CHAT AND EXTRACTION | EMBEDDINGS |
|---|---|---|---|
| OpenAI | Hosted | ✓ | ✓ |
| Anthropic | Hosted | ✓ | — |
| Mistral | Hosted, in the EU | ✓ | — |
| Ollama | Your machine or your network | ✓ | ✓ |
| Hugging Face | Hosted, or your own endpoint | ✓ | — |
| OpenAI-compatible: Azure OpenAI, vLLM, LM Studio, gateways | Wherever you run it | ✓ | ✓ |
// PUTKI
AI pipelines in production, with Putki
A pipeline that calls a model is still a pipeline: it runs on a schedule, fails at 3 a.m. and moves data someone will ask about. Putki runs it like every other.
ALL PLANS
Execution monitoring
Every run of every pipeline: what ran, what failed, rows and duration, and the log of any single run. A provider that times out shows up like any other failure.
PROFESSIONAL · ENTERPRISE
Lineage
What each run read and wrote, across projects: which datasets went into a prompt and where the answers ended up.
MIGRATIONS
Script Components, first drafted by a model
The SSIS importer can ask a local or hosted model to translate C# Script Components to Java. Every draft is compiled, checked and flagged for review; the converted pipelines need no model to run.
// LEARN MORE
Do it yourself
The Apache Hop manual, the Putki docs and our blog cover language models in pipelines from the first prompt to production.
A model that works in a notebook, but not in production yet?
We help teams turn a proof of concept into a pipeline that runs on schedule, copes with imperfect answers and is monitored, with the model you choose.
// FAQ
Things people ask
The one that fits your data, your budget and your rules. In Hop the model is a connection setting: switching from one provider to another changes the AI Provider, not the pipeline.
Only if you choose a hosted provider, and only what your pipeline puts in the prompt. With Ollama, nothing leaves your network; with Mistral, requests stay on European infrastructure. Hop sends nothing to any model until you set up a provider.
That is normal, and the pipeline handles it. Structured extract constrains the answer to the fields you define, and ordinary transforms validate, clean and route what comes back, so a bad answer becomes a rejected row, not a broken load.
No. Every AI transform and the AI Assistant are part of open-source Apache Hop. Putki adds what production needs around them: monitoring, lineage, hardened images and support.
Put your language models to work
Tell us what you want a model to do with your data, and we'll show you the pipeline.