Case study
Technicians ask in their own words. Workheld AI answers with source and page number.
Workheld AI is the AI assistant in Workheld Flow. Pondhouse Data built the AI infrastructure behind it: semantic search across the full plant documentation and agents on the new Azure AI Foundry standard, built into the Workheld platform.
- Client
- Workheld GmbH, Vienna
- Product
- Workheld AI in Workheld Flow
- Launch partner
- Swiss Post
- Scope
- AI search · Foundry agents · platform integration
- Platform
- Azure AI Foundry · Azure AI Search · Java
What the search covers
33,808
text chunks from manuals, schematics and Excel sheets. Workheld AI finds them even when the question uses completely different words than the document.
- Documents
- 1,091
- Vector index
- 403 MB
- Largest file
- up to 128 MB
- New documents
- every 2 hours
- Embedding
- 3,072 dimensions
Figures for launch partner Swiss Post. As of March 2026.
01The task
Answers from more than a thousand documents.
Workheld AI is the AI assistant in Workheld Flow, Workheld's platform for maintenance and field service. Technicians ask their question right in the web or mobile app, and the answer comes from their company's own documentation. Swiss Post was the launch partner in autumn 2025.
The documentation consists of manufacturer manuals, schematics and Excel sheets, with single files several hundred megabytes in size. The questions come in everyday language. We built the search across these documents and the agents that answer inside the platform.
Workheld AI on workheld.com02The AI search
A search that understands technicians.
The search runs on Azure AI Search. It reads every document in full, splits it into short chunks and understands what a question means. The agent only gets the matching passages and cites them as sources.
Workheld Flow · question
“Air isn't coming through anymore.”
Possible causes
Workheld AI- 01Compressor isn't building pressure
- 02Filter clogged
- 03Hose kinked or leaking
- 04Nozzle dirty
Sources
- Compressed air maintenance manual, p. 11
- Conveyor fault table, p. 4
Example with a simplified answer and made-up sources
01
Ingest
An indexer picks up new manuals, schematics and Excel sheets from storage every two hours, including files up to 128 MB.
02
Chunk
Text is split into chunks of 512 tokens with a 128-token overlap, so nothing is lost at page breaks. Each chunk gets an embedding with 3,072 dimensions.
03
Retrieve
The search combines vectors, keywords with a German analyzer for compound words, and semantic reranking.
04
Answer
The agent answers with document and page. If the documents don't cover the question, it says so.
03Agents on the Foundry standard
Built on the new agent standard of Azure AI Foundry.
Microsoft has rebuilt agents in Azure AI Foundry around versioned prompt agents, the Responses API and tools over MCP. We set up the Workheld AI agents on this standard and moved the Java service of the Workheld platform onto it.
Workheld Flow
Web and mobile
aiservice
Java · Workheld platform
Foundry agent
Prompt agent, versioned
Knowledge base · MCP
Azure AI Search
Function tools
Fault reports in the database
01
Versioned prompt agents
Each agent is defined in Foundry as a prompt agent with versions. The platform addresses it by name. Instructions can be changed and tested without redeploying the service.
02
Knowledge over MCP
The search is attached to the agent as a knowledge base over MCP. Agentic retrieval breaks complex questions into sub-queries and fetches only the matching chunks.
03
Tools for platform data
Two function tools query the fault reports in the Workheld database, for search and statistics with exact filters. That's how the agent answers “How many hydraulic faults did we have last week?”.
04
Responses API with streaming
Conversations run through the new Responses API. Answers are streamed, and sessions and chat history are kept in Workheld Flow.
05
Sources from tool calls
The service reads file name, page and link straight from the agent's search calls. So the links are right even when the model only writes “source” in its text.
06
Multi-tenant
Each Workheld customer gets its own Foundry project, index and agent. The service picks the right tenant per request. Every setup step is documented as a runbook.
04The result
What technicians can ask now.
Everyday questions asked to Workheld AI, in German. The answers come from the customer's own documentation.
“Air isn't coming through anymore”
Four causes with fixes: compressor, filter, hose, nozzle.
“Display shows nothing”
Seven causes, including fuse types and power-saving mode, with page references.
“Prepare the plant for winter”
A nine-step procedure against corrosion and condensation, from four documents.
“Lubricant for conveyor belts”
The right grease, with a warning about mixing lubricants and a link to the manual.
“Replace the drive belt”
Step-by-step instructions, plus a link to the spare parts catalogue.
≈ 50%
lower cost per question, because the agent gets short relevant chunks instead of whole documents.
2 h
until a new document can be found by Workheld AI, with no manual step.
128 MB
large manuals and schematics are read and searched in full.
Cost figure estimated during the project, not measured over a longer period.
“Our customers ask in their own words. Pondhouse built the search and the agents behind Workheld AI so that answers come from the actual documentation, with sources, and on the current Azure AI Foundry standard.”
05Who it's for
When we can help.
Good fit
- 01You build software and want AI built into your product.
- 02Answers should come from manuals, tickets or databases, with sources.
- 03You want to run agents on Azure AI Foundry in production, not just try them out.
Probably not a fit
- 01You're looking for an off-the-shelf chatbot.
- 02There are no documents or data the answers could rely on.
- 03You need a prototype for a presentation that will never go live.
Next step
Want AI search and agents in your software?
Tell us in 30 minutes what your assistant should do and where the answers should come from. We'll tell you how we would build it.