Oman Power Plant Telemetry Intelligence Engine with LlamaIndex & Deepgram
Managing a $683m CCGT power plant contract in Oman requires querying complex turbine manuals while processing voice updates from site engineers. Build a voice-enabled RAG pipeline using LlamaIndex for indexing plant engineering docs and Deepgram for real-time acoustic transcription of engineer status reports. Achieve an exact operational retrieval recall above 90%.
What you are building
The core problem, expected build, and operating context for this challenge.
Build an audio-driven power plant inspection RAG assistant using LlamaIndex indexing and Deepgram transcription.
How work is evaluated
Evaluates audio-to-RAG accuracy on CCGT plant maintenance queries.
Challenge datasets
Public datasets, plus the private datasets you have access to.
How submissions are scored
These dimensions define what the evaluator checks and which criteria separate a passable run from a strong one.
Transcription Accuracy
Ensures transcript contains key domain terms like 'casing B' and 'hot start'.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
Retrieval Recall@3
Percentage of queries where correct manual section is in top 3 retrieved results. • target: 0.9 • range: 0-1
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
What you should walk away with
Transcribe site voice memos using Deepgram Speech-to-Text SDK
Index thermal power plant documentation with LlamaIndex vector indices
Combine audio queries with semantic search over engineering documents
Deliver actionable maintenance summaries to plant engineering managers
Reference links and supporting material
A technical manual dataset covering 1,000 pages of CCGT power plant specs and 50 recorded voice queries from Oman site engineers.
Challenge setup
Evaluates audio-to-RAG accuracy on CCGT plant maintenance queries.
Challenge input
Audio file path or base64 WAV stream containing technical question
LlamaIndex
Framework for enterprise document indexing.
Deepgram
Speech-to-text engine for site operator audio feeds.
Evaluated output
JSON object with transcript, document_chunk_id, and final answer
- Ensures transcript contains key domain terms like 'casing B' and 'hot start'.
- Percentage of queries where correct manual section is in top 3 retrieved results. • target: 0.9 • range: 0-1
- Benchmark: RAGBench-Audio
- Retrieval Recall@3 target: 0.9
- 1 public reference case
- Python execution harness
- Python sandbox (unavailable on Versalist)
Summary shows the first task, up to three tools, and up to three evidence items. Open the technical recipe to view all configured items.
View technical recipe
Configured tools
- Deepgram · Required
- Llama Index · Optional
- Exa · Optional
Evaluation contract
- Transcription Accuracy · Weight 1
- Retrieval Recall@3 · Weight 1
Metrics
- Retrieval Recall@3 · Target 0.9
Benchmarks
- RAGBench-Audio
- general
Reference cases
- 1 public reference case
Workflow tasks
- Task 1
Input: Audio file path or base64 WAV stream containing technical question
Output: JSON object with transcript, document_chunk_id, and final answer
Execution harness: Python
Evaluation tools
- LlamaIndex · Data Orchestration / RAG
- Deepgram · Audio Speech Recognition
Recipe freeze marker
No recipe freeze time is recorded.