Deterministic troubleshooting system for aviation ground support equipment.
RAGOps is a Retrieval Augmented Generation pipeline built for GSE maintenance teams. It ingests OEM technical manuals, indexes them locally, and answers troubleshooting questions using cited source text. If the documentation does not support a claim, the system refuses.
No cloud dependency. No API keys. No data leaves your infrastructure.
- As of now, the whole E2E and retrieval pipeline is complete (hybrid search, phrase-match bonus, domain boosting[topic labeling => a *"closed envelope"])
- Initial manual testing: correct procedure identification on JETEX 4D queries
- Benchmark harness designed; corpus build in progress (50-100 Q&A pairs from GSE manuals)
- Youtube video where I'm going over:
- result discussions, distinctions and embedding strategies now and the what's coming up (training a local SLM? Yes.)
- why another RAG pipeline? What provides distinction with this approach (beyong the hybrid approach and phrase matching)
- walk through architecture stages E2E and discuss a bit why I chose "this over that"
- Demo full pipeline E2E
- Talk a bit about my background why GSE, why GPUs and *"closed envelope"
A technician is on the ramp. A GPU does not build voltage. The manual is 170 pages. The experienced technician who knew the unit retired. The new technician searches online and finds a forum post from 2011.
Two hours later the issue remains unresolved.
This repeats across fleets, shifts, and new hires.
Current options:
- Paper manuals that are slow to navigate
- Generic AI systems that invent specifications
- OEM support contracts with long response times
In aviation maintenance, an incorrect specification is worse than no answer.
RAGOps converts OEM maintenance manuals into a private troubleshooting system.
It provides:
- Page citations
- Exact specifications extracted from source text
- Symptom to procedure alignment
- Explicit refusal when context is insufficient
It does not guess.
Supported documentation can include Hobart, TLD, Guinault, ITW, and internal fleet manuals.
Tested against the Hobart JETEX 4D Spec 7003B operations and maintenance manual.
- 170+ pages
- 79 indexed chunks
- Approximately 40,000 tokens
Question:
The generator does not build voltage when operating the push to build up switch. What resistance reading should be measured on the revolving field, and which wires must be disconnected?
Result:
- 10 to 11 ohms
- Disconnect yellow and red orange wires to the generator
Verified against Chapter 3-2. Exact match. Cited.
Question:
The engine starts but shuts down when switching from START to RUN. To test overtemperature protection, where is the jumper placed and what temperature does the switch trip?
Result:
- Place jumper on water temperature switch terminals
- Correctly identified S402 procedure
- Refused to provide trip temperature because it was not present in retrieved context
If a value is not present in the source material, it is not generated.
Fleet operators maintain internal documentation:
- Modified wiring diagrams
- Maintenance schedules
- Incident reports
- Compliance records
- Airport specific procedures
Sending this data to external APIs introduces governance risk.
RAGOps runs fully on local hardware:
- Llama 3.1 via Ollama
- BGE embeddings locally
- No outbound network requirement
- Can operate without internet access
For military, government, or regulated operators, this is required.
- Parse PDF
- Sliding window chunking at 400 tokens with 25 percent overlap
- Generate embeddings locally
- Maintain page traceability
- Create versioned index files
- No modification after promotion
- Reproducible retrieval state
Before activation:
- Empty retrieval checks
- Confidence threshold checks
- Hit count validation
- Noise detection
Failing indexes are not promoted.
Hybrid retrieval:
- Semantic similarity using BGE
- Keyword scoring using BM25
- Reciprocal Rank Fusion merge
- Phrase bonus for multi word matches
- Domain score boost
Top ranked chunks are passed to inference.
Model instructions enforce:
- Use only provided context
- Match symptoms to correct procedure
- Refuse unsupported claims
Responses include chunk citations.
Each query records:
- Retrieved chunks
- Confidence scores
- Domain detection
- Grounding sources
Full traceability.
PDF → hash + extract → chunk (400 tokens, 25 percent overlap)
→ embeddings → versioned index → evaluation gate → promotion
Query → embed
→ semantic + keyword retrieval
→ rank merge
→ phrase bonus
→ domain boost
→ top chunks
→ grounded inference
→ cited answer or refusal
- Deterministic retrieval
- Immutable index versions
- Evaluation required before activation
- Explicit refusal when unsupported
- Configurable local model backend
No approximate nearest neighbor search. No stochastic retrieval layer.
- Embeddings: BGE base en v1.5
- Model: Llama 3.1 8B via Ollama
- Search: Cosine similarity plus BM25
- Storage: JSON versioned indexes
- Tests: 119 passing
docker compose up -d --build
docker compose exec ragops bash
ragops run --pdf /app/data/test_pdfs/JETEX4D7003B.pdf
First model load may take one to two minutes. Later queries typically complete within 15 to 30 seconds depending on hardware.
git clone https://github.com/ereztdev/RAGOps.git
cd RAGOps
pip install -e .
ragops run --pdf path/to/manual.pdf
Step by step:
ragops ingest --pdf path/to/manual.pdf
ragops build-index --index-version v1 --overwrite
ragops ask --query "Your troubleshooting question"
| Command | Purpose |
|---|---|
| ingest | Parse PDF into chunks |
| build-index | Create embedding index |
| evaluate | Run retrieval checks |
| promote | Activate evaluated index |
| ask | Query active index |
| run | End to end workflow |
- Python 3.10 or higher
- sentence transformers
- pypdf
- Ollama
AGPL 3.0