How to add an AI chatbot to proALPHA
Short answer
Build the chatbot as a separate web or Teams-style front end that answers from proALPHA data through approved tools and document retrieval, respecting user roles. It can run completely on-prem with an open-weight model.
A chatbot for proALPHA is a different thing from an agent. The goal is simple: a person asks in plain language, in German or English, and gets an answer from the ERP and from company documents, with sources. Staff want it for status and how-to questions. Some companies also want a customer or supplier portal assistant.
proALPHA's screens are powerful but dense, and new users take time to learn where things live. A chatbot shortens that path: where is the stock of this item, who is the supplier for it, what is the status of this production order, how do I book a scrap quantity in our process.
The decision that matters is the audience. An internal chatbot can reuse broad staff access. A customer-facing one must be tightly scoped to that customer's own orders, with a much stricter data boundary.
What teams use it for
Internal status assistant
Staff ask about order, production and stock status without navigating menus. Answers show the numbers and the source record.
How-do-I help for new users
The bot answers process questions from your own work instructions and proALPHA documentation, pointing to the relevant screen by name.
Customer order portal assistant
A customer asks about their order and gets status, dates and documents, restricted to their own account.
Supplier self-service
Suppliers ask which of their PO lines are open and what dates are expected, and submit a new confirmed date for review.
Shop floor question line
Operators ask about routing steps, tooling and work instructions from a tablet, with answers drawn from controlled documents.
proALPHA integration surface
- OpenEdge database or exposed services supply live answers through a tool layer. Use read-only access, and prefer proALPHA services where available so rights logic is reused.
- Documents: work instructions, quality manuals and drawings usually live in file shares or a document management system, and are indexed separately for retrieval.
- Identity: map the chatbot user to a proALPHA user or role. Active Directory or Entra ID single sign-on is typical for staff. Customers need a separate identity store.
- Portal and web front ends: proALPHA customers often add web portals; the chatbot can be embedded as a widget in an existing intranet or portal.
- Notifications: email or a messaging channel can carry follow-ups when the bot cannot answer.
How it works
- 1
Pick audience and scope
Start internal and read-only. Decide which topics the bot answers and which it refuses.
- 2
Index the documents
Chunk work instructions and manuals, keep their access rights as metadata and refresh nightly. Include German and English versions.
- 3
Build the data tools
Create tools for orders, stock, items and purchasing, each returning small structured results.
- 4
Add routing
Decide per question whether to use documents, data tools or both. Keep a fallback that says what is not known.
- 5
Enforce identity and rights
Pass the verified user identity into tools and the retrieval filter, so the bot cannot reveal more than the user could see.
- 6
Run a pilot with feedback buttons
Give twenty users access, collect thumbs-down reasons and fix the worst failure patterns weekly.
- 7
Expand cautiously
Add new topics one at a time. Consider customer access only after the internal version is stable.
Example flows
Operator asks about a work step
Trigger: Tablet question: what torque for step 30 on this assembly
- Bot identifies the item and routing from the production order context.
- It retrieves the controlled work instruction for that revision.
- It answers with the value and shows the document name and revision.
Result: The operator gets a sourced answer without leaving the workstation.
Customer asks where their order is
Trigger: Customer logs into the portal and types a question
- Bot authenticates the customer account.
- Tool returns only that customer's open order lines.
- Bot answers with promised dates and offers to notify them of changes.
Result: Fewer status emails to customer service, with no cross-customer data exposure.
Keeping data on-prem
A proALPHA chatbot can be fully on-prem: an open-weight model served with vLLM or Ollama on your GPUs, a local vector index for documents, a tool service reading OpenEdge on a replica, and a web front end on your intranet. For German-language use, choose and test models that handle German technical text well, since quality varies. Customer-facing bots are harder on-prem because the front end must be reachable from outside; place only a thin gateway in the DMZ and keep the model and data internal.
Security checklist
- Role-aware retrieval and tools: the bot sees only what the signed-in user may see.
- Strict separation between internal and customer-facing instances, with different tools and indexes.
- Refuse topics outside scope, and never answer from model memory about prices or contract terms.
- Filter and rate-limit external traffic to prevent scraping through the chat interface.
- Retain chat logs under a stated policy, since they can contain personal data.
Ways to get it done
Use the vendor or hosting partner's assistant features
Effort: Low, but verify scope and data handling.
Enough if your release includes a suitable assistant and you only need help with standard screens.
Build on an open-source chat and RAG stack
Effort: Several weeks for an internal pilot.
Teams with web and ML engineering capacity and a documents-first need.
Netray-built private assistant
Effort: Pilot in a few weeks, then widen topics.
Companies wanting live ERP answers plus documents, on-prem, with an evaluation set.
Pitfalls
- !Indexing documents without revision control, so the bot quotes an obsolete instruction.
- !Letting the bot answer price questions from memory instead of data.
- !No identity mapping, so everyone sees everything.
- !Poor German handling because testing used only English prompts.
- !No feedback loop, so early errors destroy trust.
Frequently asked questions
Can the chatbot show live proALPHA data?
Yes, through tools that query OpenEdge on a replica or call exposed proALPHA services. The bot answers from the returned rows and cites them, rather than from the model's memory.
Does it work in German?
It can, but model quality differs. Test with real German terms from your own master data and work instructions, including abbreviations. Pick the model after that test, not before.
Can customers use it?
Yes, with a separate, tightly scoped deployment: customer identity, only their own orders, no internal documents. Build and prove the internal version first.
How is this different from the ChatGPT or Claude options?
Those are general assistants you connect to tools. A chatbot here is a managed front end for your users, with your branding, roles, documents and logging, and it can run entirely on your own hardware.
What hardware do we need on-prem?
For a pilot, one server with a modern GPU is often enough for a mid-sized open-weight model and a few dozen users. Document indexing is lighter. Size it after measuring your real question volume and response-time target.
Want AI chatbot working on your proALPHA?
A free 30 minute assessment: we map the integration path for your version and setup, and where it should run.
Get a free assessment