AI-native roles
Hire RAG Engineers
Retrieval specialists who keep answers grounded when your corpus grows from a thousand documents to ten million.
A RAG engineer builds retrieval-augmented generation systems: chunking, embedding, hybrid search, reranking and grounded evaluation. Hire one when your AI answers are confidently wrong or degrade as the corpus grows. Expect $85 to $160 per hour depending on scale and evaluation rigour.
What this role actually does
Retrieval-augmented generation looks simple in a tutorial: embed documents, search by similarity, paste results into a prompt. It stops being simple at roughly the point where the corpus exceeds what a single engineer can eyeball. Chunk boundaries start splitting the sentence that held the answer. Similarity search returns documents that are topically close and factually irrelevant. Recall collapses quietly, and nobody notices because the model still produces a fluent answer.
A RAG engineer treats retrieval as an information-retrieval problem rather than a prompting problem. That means measuring recall and precision against a labelled set, using hybrid lexical-plus-vector search rather than vectors alone, reranking candidates before they reach the model, and building an evaluation harness that catches groundedness regressions before deployment.
Core skills to look for
- Chunking strategy: semantic, structural and hierarchical, chosen per document type
- Embedding model selection and honest benchmarking on your own corpus
- Hybrid retrieval: BM25 or full-text combined with dense vectors
- Reranking with cross-encoders and the latency budget to afford it
- Vector stores: pgvector, Pinecone, Weaviate, Elasticsearch, and their real trade-offs
- Groundedness and faithfulness evaluation, plus citation verification
- Query understanding: rewriting, decomposition and multi-hop retrieval
- Freshness and incremental indexing without full rebuilds
When you need to hire this role
- Your assistant gives confident answers that are not supported by your documents
- Quality was fine at ten thousand documents and fell apart at a million
- You cannot measure whether a retrieval change helped or hurt
- Users complain that the system misses documents they know exist
- Latency is unacceptable once reranking is added naively
How to assess candidates for this role
Ask how they measure recall
A candidate who cannot describe building a labelled evaluation set has not operated RAG at scale. Listen for recall@k, mean reciprocal rank and groundedness scoring.
Probe chunking judgement
Ask how they would chunk a 200-page contract versus a support knowledge base. Different answers for different document structures indicate real experience.
Test hybrid search reasoning
Ask when pure vector search fails. Strong answers mention exact identifiers, product codes, rare terms and negation — cases where lexical matching wins outright.
Check the latency trade-off
Ask how they would add reranking without breaking a 500ms budget. Expect candidate-set trimming, cascade ranking and caching rather than a shrug.
Typical rate bands
| Level | Experience | Typical rate | What they own |
|---|---|---|---|
| Mid-level | 4-6 years | $85-115/hr | Improves an existing pipeline, owns chunking and indexing work |
| Senior | 7-10 years | $115-145/hr | Owns retrieval architecture and the evaluation harness end to end |
| Staff / Principal | 11+ years | $145-160/hr | Designs multi-corpus retrieval platforms and org-wide evaluation standards |
Frequently asked questions
What does a RAG engineer do?
They build and tune retrieval-augmented generation systems: choosing chunking strategies, selecting and benchmarking embedding models, combining lexical and vector search, reranking results, and building evaluation harnesses that measure groundedness rather than relying on subjective spot checks.
Why do RAG systems get worse as the corpus grows?
Because similarity search returns a fixed number of results from a growing pool, so the chance that the right document appears in the top results falls. Without reranking, hybrid search and measured recall, quality degrades quietly while the model keeps producing fluent answers.
Do I need a RAG engineer or an AI agent engineer?
If your problem is answer accuracy over a document corpus, hire a RAG engineer. If your problem is a model taking actions through tools, hire an AI agent engineer. Many production systems eventually need both, but the assessments are genuinely different.
How much do RAG engineers cost?
Typical vetted contract rates are $85 to $115 per hour at mid-level, $115 to $145 at senior and $145 to $160 for staff and principal engineers, varying with corpus scale, latency requirements and evaluation rigour.
What is the single most common RAG mistake?
Relying on pure vector similarity with no lexical component and no reranking. It works acceptably in a demo and fails badly on exact identifiers, rare terminology and negation, which are exactly the queries users care most about.