RAG pipeline development — ingest, embed, store, retrieve, re-rank, generate
Production-grade RAG with evals. Chunking strategy, embedding choice, re-ranking, eval harness — all wired correctly.
Who this is for
Company whose knowledge base or documents would be more useful if an LLM could actually answer questions against them.
The pain today
- First-try RAG is hallucination-prone.
- Chunking strategy picked at random, never tuned.
- Embedding choice made on a blog post, not measured.
- No re-ranking, so retrieval quality is 'cosine similarity and hope'.
The outcome you get
- Production RAG pipeline: ingest, embed, store, retrieve, re-rank, generate.
- Chunking strategy tuned with an eval set.
- Embedding model picked by measurement, not hype.
- Re-ranking layer improving retrieval quality 30 to 50%.
- Eval harness (RAGAS or custom) catching regressions.
The full RAG pipeline
A production RAG pipeline has six stages: (1) Ingestion — pull documents from source (Notion, Google Drive, S3, database), normalize to text plus metadata. (2) Chunking — split into 512 to 1024 token chunks with 10 to 20% overlap, tuned per document type. (3) Embedding — OpenAI text-embedding-3-large as default, Cohere embed-v3 for multilingual. (4) Storage — pgvector under 10M vectors, Pinecone or Weaviate above. (5) Retrieval plus re-ranking — top-k cosine search plus Cohere Rerank or cross-encoder (30 to 50% quality improvement). (6) Generation — LLM call with retrieved context plus output validation. Most teams ship half.
Instill as the AI-product reference
Instill (SITE-FACTS §6) is an AI skills platform running Next.js 16 plus React 19 plus PostgreSQL plus Vercel plus MCP Protocol. 1,000+ skills saved, 45+ projects powered, 30+ active users. The retrieval patterns used in Instill (semantic search across skills, context-aware retrieval for agents) transfer directly to customer-facing RAG. If you are adding a 'chat with your docs' or 'answer questions over our knowledge base' feature, the pattern library is proven.
Vector store choice, honestly
The vector store pick depends on scale. Under 10M vectors: pgvector inside your existing Postgres (no extra vendor, no extra bill, no new operational surface). 10M to 100M vectors: Pinecone for managed convenience or Weaviate self-hosted for cost ceiling. Above 100M: serious evaluation required. Most RAG projects live below 10M vectors, which means most RAG projects should use pgvector and never bring in a dedicated vector DB. The engagement measures before committing.
Pricing and scope
AI Automation retainer at $3,000 per month. Typical RAG engagement: 6 to 12 weeks for the first production pipeline (ingestion plus embedding plus retrieval plus eval harness), then ongoing tuning under the retainer. 14-day money-back. Cancel anytime.
Recent proof
A comparable engagement, delivered and documented.
A prompt library that works with every AI tool
A home for your best AI prompts. Save them once, then use them in Claude, Cursor, or any AI tool you work with. No more copy-paste.
Frequently asked questions
The questions prospects ask before they book.
- pgvector, Pinecone, or Weaviate?
- pgvector under 10M vectors. Pinecone for managed convenience at 10M to 100M. Weaviate self-hosted for cost ceiling at the same scale. Above 100M — real evaluation required.
- Chunking strategy?
- Start with 512 to 1024 tokens with 10 to 20% overlap. Semantic chunking (based on document structure) beats fixed chunks for most document types. Tune with eval set.
- Embedding model?
- OpenAI text-embedding-3-large as default (strong, cheap, available). Cohere embed-v3 for multilingual. Open-source (BGE, E5) for self-hosted or cost-ceiling requirements.
- Re-ranking — is it worth the latency?
- Almost always yes. Cohere Rerank adds ~100ms and typically improves answer quality 30 to 50%. Cross-encoder models on HuggingFace work too for self-hosted.
- Eval harness — what metrics?
- RAGAS: faithfulness (is the answer grounded in retrieved context), answer-relevance (does the answer address the question), context-relevance (is the retrieved context actually relevant). Custom metrics on top when the domain requires.
Ready to start?
Tell me what you need in 60 seconds. Tailored proposal in your inbox within 6 hours.