You shipped a RAG pipeline. Answers are hit-or-miss. You want someone who's done this before to walk through your setup and call the shots — embedding model, chunk size, retrieval strategy, eval harness — and tell you what to fix first.
Token cap vs your real chunk distribution (Nomic capping at 2048 silently, OpenAI ada-002 vs text-embedding-3-large vs the open-source alternatives), domain fit, dimension/storage trade-offs.
Token-vs-character, semantic vs fixed-size, overlap, table/code handling. The chunker is the biggest unforced error in most RAG pipelines.
Vector-only vs hybrid (BM25 + vector + rerank), top-k tuning, metadata filtering. Whether you actually need a reranker or whether your top-1 is fine.
pgvector vs Pinecone vs Qdrant vs Azure AI Search. HNSW parameters. Index type for your scale. When you should switch and when you should stop fiddling.
Whether you have one. If not, the lightest-weight one that catches regressions: a question set, ground-truth chunks, recall@k. If yes, whether it's measuring what you actually care about.
Prompt template, citation enforcement, refusal handling when retrieval is weak, hallucination guardrails.
Each item names the symptom, the cause, the fix, and the rough effort. Sent within 24 hours of the call.
A minimum eval set sized to your domain — question count, ground-truth shape, what to measure first.