Use case: Multi-modal service co-pilot.
Technical architecture for field service: shorter mean time to repair through context-aware solution knowledge, right at the machine.
30 minutes, online, free and non-binding.
A RAG co-pilot searches maintenance logs, manuals and wiring diagrams. Hybrid search finds exact error codes and semantic problem descriptions at the same time. A photo of the nameplate plus a voice message automatically filter for the right device, and every answer comes with a cited source. Runs on-premise or in a private cloud.
This is the technical blueprint for a solution we build for field service technicians. Performance figures will only be published once measured in a concrete project.
The scenario: a technician is standing in front of a machine that's down, has no internet signal in the basement, and needs access to thousands of pages of manuals, wiring diagrams and maintenance logs.
We're not building a simple search engine here, but a retrieval-augmented generation system (RAG) with hybrid search. The technical goal: shorter repair times through context-aware solution knowledge in a matter of seconds. And just as important: avoiding hallucinations as far as possible through strict grounding.
Three layers of knowledge processing
Preparing documents
Unstructured data (PDFs, TIFF wiring diagrams, Word) gets processed via OCR. For parsing and splitting into manageable chunks, we use tools such as Unstructured.io.
- Example: RecursiveCharacterTextSplitter
- Chunk size 1000 tokens, overlap 200
Storing meaning
We store the semantic meaning of the chunks as embeddings (vectors), for example in Qdrant or Pinecone, for fast similarity search.
- Embedding model depends on hosting variant
- Metric: cosine similarity
Generating the answer
The orchestrator (for example LangChain) connects the question with the retrieved knowledge and generates the answer. The ReAct pattern matters here for logical reasoning.
- Strict grounding
- Mandatory sourcing
Why vector search alone isn't enough
- Understands concepts like "the motor is running hot"
- Finds "hydraulic overheating issues" even if someone just types "the pump is warm"
- Often fails on exact identifiers like "error code E-404-X"
- BM25 searches exact keywords like part number "XJ-500-2" or error code "E33"
- Vector search adds the semantic connection
- Reciprocal Rank Fusion (RRF) merges both result lists and passes the best result to the language model
Four steps on site
Input processing
The technician photographs the nameplate and dictates: "What's the torque spec for the main bolt on this series?" Multimodal input from image and speech.
Query transformation
The system turns the often vague question into precise search queries (query expansion). Metadata filters, driven by image recognition, automatically narrow the search to the right machine series.
Citation & verification
The language model generates the answer, for example the prescribed torque value for a bolted connection, with a link to the original PDF and page number. No verifiable source, no output. That's our hard guardrail.
Feedback loop
The technician clicks "worked" or "wrong." This feedback feeds back into improving the index and the answers.
What carries the business case
Response time
Target: an answer in a matter of seconds, even for complex hybrid searches. Measured during pilot operation.
Context window
Depends on the chosen model and gets fixed during pilot operation.
On-premise
Private cloud or on-site server, your choice. No training on public data.
Fewer follow-up queries
Target: fewer escalated follow-up queries in the back office and first-level support, measured during pilot operation.
The RAG architecture check for industry and service
Data ingestion setup
The blueprint for turning your own manuals into a working AI service.
Embedding and chunking
Best practices for preparing technical documents.
Tech FAQ
What's the difference between vector search and hybrid search?
Pure vector search finds documents by semantic connection. That's good, but often not enough in technical contexts. Hybrid search combines it with BM25 keyword matching. That way you find both the conceptual solution and the exact part number that a vector search easily misses.
Does the service co-pilot also work without an internet connection?
The core architecture can run entirely in your private cloud or on on-premise servers. If the technician is in the basement with no signal, the request can be buffered until the company network or mobile signal returns, depending on the deployment strategy.
How are AI hallucinations prevented?
We minimize hallucinations through strict grounding in the orchestrator. The model refers exclusively to the retrieved document chunks. If the manual contains no information, the answer is "information not found." In addition, every statement is linked with a source and page number.
Which document formats can be processed?
The ingestion layer processes nearly all unstructured formats: text files, Word documents, complex PDFs (including multi-column), HTML, and, via OCR, image files such as scanned TIFF wiring diagrams.
Shall we build your technical co-pilot?
The technology is ready. The challenge lies in clean data preparation and the architecture. Let's check whether your data is ready for RAG.

