RAG consulting · Retrieval · Grounding · Evaluation

RAG systems that answer from your data and show their sources.

We design, build and evaluate retrieval-augmented generation (RAG) systems: ingestion, retrieval, grounding, evaluation and security. Led by Manmeet Singh, founder of AIshar Labs.

AIshar Labs designs, builds and evaluates retrieval-augmented generation (RAG) systems: ingestion and chunking, embeddings, search and reranking, grounding and source attribution, and evaluation. We have worked with Elasticsearch, Lucene, two-tower retrieval models and RAG pipelines, in prior roles and in consulting work, and we build on Google Cloud and AWS. A RAG Quality Audit is available by request. Work runs as an Architecture Sprint, an Embedded Build or a Targeted Build.

Your RAG answers need to be right, and provably so.

Five parts of a RAG system that holds up.

A RAG answer can only be as good as what gets retrieved, so most of the work is before the model.

Part 01

Ingestion and chunking

How your documents and data are cleaned, split and prepared so the right passages can be found later.

Part 02

Retrieval and reranking

Embeddings, vector search, keyword and hybrid retrieval, metadata filters and reranking. We have worked with Elasticsearch, Lucene and two-tower retrieval models.

Part 03

Grounding and source attribution

Answers tied to their sources, with citations a user or auditor can check. The question is whether an answer is grounded, or filled in by the model.

Part 04

Evaluation and cost

Measuring retrieval precision against context size to find the quality and cost frontier, and building the evaluation habits that show whether a change helped. RAG systems often send 10 to 20 chunks when 3 would do.

Part 05

Security and compliance

Tenant isolation in retrieval, PHI and PII handling, logging and audit trails. In our healthcare reference build, PHI was de-identified before embedding and every answer was traceable to its sources.

Delivered through the engagement models you already know.

RAG consulting is an area of expertise, not a separate product, so there is no separate price list. The same models apply as on How We Work.

RAG Quality Audit, by request: a deep diagnosis of retrieval, grounding, citation accuracy and evaluation maturity for teams whose RAG system is giving inconsistent or unverifiable answers. Related: HIPAA AI infrastructure, search relevance consulting, LLM cost optimization and the fine-tuning vs prompting vs RAG decision framework.

An online clothing retailer's recommendations were basic or wrong. We built a system where shoppers type or speak, a search agent or voice agent calls RAG over the retailer's own data, Gemini models handle classification and decisions, and OpenAI models serve as the frontier models, all on GCP with Terraform. Click-through rate improved by 15 to 20 percent and add-to-cart by 5 to 7 percent.

Read the full story →

Questions teams ask before they book.

What is RAG?
Retrieval-augmented generation. The system retrieves relevant passages from your own data at question time and gives them to a language model, so the answer can draw on your sources instead of only on what the model learned in training.
How is RAG different from fine-tuning?
RAG supplies relevant information at answer time. Fine-tuning changes the model itself. Which fits depends on the task, and we cover the decision in our article on fine-tuning versus prompt engineering versus RAG.
What is a RAG quality audit?
A deep diagnosis of retrieval, grounding, citation accuracy and evaluation maturity, for teams whose RAG system gives inconsistent or unverifiable answers. It is available by request.
How do you evaluate a RAG system?
We look at whether the right passages are retrieved, whether answers are grounded in them, whether citations are accurate, and how mature your evaluation is. We also benchmark retrieval precision against context size to find the quality and cost frontier.
Can RAG be used with sensitive or regulated data?
Yes, with the right architecture. In our healthcare reference build, PHI was de-identified before embedding, tenants were isolated at the index level, and every answer was traceable to its sources. See HIPAA AI Infrastructure. Compliance decisions stay with you and your counsel.
Which technologies do you work with?
We have worked with Elasticsearch, Lucene, two-tower retrieval models and RAG pipelines, in prior roles and in consulting work, and we build on Google Cloud and AWS with language model APIs such as OpenAI and Gemini and with open-source models.
Who owns the work?
You do. You receive the complete codebase and documentation, with zero lock-in. See How We Work for details.

Tell us what you're building. We'll tell you honestly if we can help.

Book a call
Manmeet Singh
Manmeet SinghFounder · ML Architect