Retrieval-Augmented Generation (RAG)
A method that retrieves relevant knowledge before generating an answer, supporting traceable responses and document updates without retraining the model.
Definition
Retrieval-augmented generation first finds relevant document passages, then gives them and the question to a language model. This supplies private enterprise knowledge and information newer than model training. Updating the external knowledge base can update answers without retraining, and source references allow users to check the basis of a response.
Basic Workflow
Offline processing divides documents into meaningful chunks, computes embeddings and builds indexes. Online processing interprets or rewrites the question, retrieves candidates, reranks them and generates a cited answer. Tiny chunks lose context; oversized ones dilute relevant details. Missing or incorrect retrieval limits answer quality.
Retrieval Quality
Hybrid search combines semantic similarity with exact keyword matching for model names and identifiers. Query rewriting handles informal language or splits complex requests. Reranking improves candidate order. Metadata filters restrict equipment, applicability and document version so superficially similar but unsuitable material is excluded.
Industrial Challenges
Tables, drawings, exploded views and scans need careful parsing; separated headers and values can mislead. Manuals vary by year and configuration. Synonyms differ between manuals, work orders and spoken terms. Incomplete repair records add noise. Knowledge governance is therefore a major part of industrial RAG.
RAG and Fine-tuning
RAG supplies updateable facts with traceable evidence. Fine-tuning shapes response patterns, terminology and task conventions. For equipment parameters, start with retrieval; for consistent output behaviour, prompts or fine-tuning may help. Encoding large changing fact sets through fine-tuning is costly to maintain and correct.
Evaluation and Improvement
Maintain real questions with reference answers and expected documents. Test retrieval coverage and ranking separately from answer accuracy and faithfulness after changes to chunking, search or prompts. Classify user-reported errors as missing retrieval, faulty source material or unsupported generation, then fix the responsible layer.
Frequently Asked Questions
Should I choose RAG or fine-tuning?
Use retrieval for private or changing facts, and fine-tuning or prompts for behaviour and format. They are complementary rather than interchangeable.
Why can a RAG answer be irrelevant?
Retrieval may have missed the right material, split a table incorrectly or selected the wrong manual version. Inspect sources and ranking before assuming the generator is the cause.
Related Terms
From Concept to Fleet Deployment
AI-Driven provides ADAS perception, smart onboard terminals and fleet management platforms for industrial vehicles, turning these concepts into real gains in fleet safety, efficiency and maintenance.
This glossary covers general industry concepts for technical reference and knowledge sharing. It contains no proprietary information from any specific manufacturer. For equipment specifications and operating procedures, always refer to the OEM documentation.