Question 1

How do you stop a RAG system from hallucinating?

Accepted Answer

Three layers. First, retrieval quality: hybrid search and reranking so the model actually receives the right context instead of a loosely related passage. Second, grounding: every answer cites the source it came from, and answers that cannot be grounded are refused rather than invented. Third, evaluation: an automated harness measures faithfulness and retrieval accuracy so regressions are caught before release, not by your users.

Question 2

pgvector or Pinecone, which should we use?

Accepted Answer

It depends on scale, existing infrastructure, and team. If you already run Postgres and your corpus is in the low millions of chunks, pgvector usually wins on simplicity and cost. Dedicated vector databases earn their keep at larger scale or with heavy metadata filtering and high query concurrency. The honest answer comes from your numbers, and it is a decision made in the assessment rather than by default.

Question 3

How long does a production RAG system take to build?

Accepted Answer

The documented range across shipped systems is 4 to 8 weeks from kickoff to production. A working proof of concept on your own data usually lands within the first week, because the codebase and data are absorbed in 24 to 48 hours rather than through months of discovery.

Question 4

Can this run inside our own cloud for compliance?

Accepted Answer

Yes. Where requirements demand it, the whole system runs inside your cloud on your own accounts and keys, and your data never leaves your environment. Retrieval systems here have shipped into HIPAA-compliant healthcare under a signed BAA, and into environments built to SOC 2 and ISO 27001 requirements.

Question 5

Do you use our documents to train a model?

Accepted Answer

No. Your data is used only to build and evaluate your system. Nothing is used to train models, and the commercial API tiers these systems run on contractually exclude your data from provider training.

	Typical agency or DIY	SASID
Retrieval quality	Cosine similarity on one embedding model	Hybrid search plus reranking, tuned on your data
Hallucinations	Hope the prompt handles it	Citation grounding plus hallucination detection
Proof it works	A demo on cherry-picked questions	An eval harness with retrieval and answer metrics
Who builds it	Junior team learning on your budget	One senior engineer, start to production
After launch	Handoff and goodbye	90 days of monitoring and tuning included
Ownership	Vendor lock-in, black box	You own all code, prompts, and evals

RAG Consulting That Ships to Production.

What a production RAG build includes

Retrieval that actually retrieves

Answers you can defend

Evaluation and operations

Proof from shipped retrieval systems

RAG consultant vs the alternatives

Three ways to start

Proof-of-concept sprint

Production build

Fractional advisory

What it costs

RAG consulting questions, answered directly

See your RAG system answer correctly before you commit.