MS-RAGS(ALL-IN-ONE) Docs

Decision guide

Recommendations

Use these recipes when users are unsure which RAG, embedding, database, or retrieval option to choose.

Recommended Stacks

Beginner

Local demo

RAG: Naive RAG. Embeddings: OpenAI small or Hugging Face all-MiniLM. DB: ChromaDB. Retrieval: Dense vector. Add nothing else until the baseline works.

Production default

Support assistant

RAG: Advanced RAG. Embeddings: OpenAI small/large. DB: Qdrant or Pinecone. Retrieval: Hybrid or Dense + reranker. Evaluation: RAGAS + LangSmith/Langfuse.

Privacy

Local/private deployment

RAG: Naive or Parent-Child. LLM: Ollama local. Embeddings: BGE, E5, or nomic-embed-text. DB: FAISS, ChromaDB, Qdrant local, or PGVector.

Enterprise

Cloud enterprise

RAG: Advanced or Adaptive. LLM: Azure OpenAI or AWS Bedrock. DB: Azure AI Search, OpenSearch, PGVector, or MongoDB Atlas depending on existing stack.

Long docs

Legal/research documents

RAG: Parent-Child or Contextual Compression. Chunking: Sentence, paragraph, or document-aware. Retrieval: MMR or parent-child. Reranking: BGE/Cohere.

High quality

Strict grounded answers

RAG: Self-RAG or Corrective RAG. Enhancement: query rewriting. Retrieval: hybrid + reranking. Evaluation: CI/CD thresholds for faithfulness and context precision with tools like RAGAS or DeepEval.

Decision Matrix

User priorityRecommended choiceWhy
Lowest setup frictionOpenAI + ChromaDB + Recursive Character + Dense VectorFew moving parts and predictable behavior.
No local model downloadsHosted OpenAI/Cohere/Hugging Face endpoint embeddingsAll embedding inference happens through API calls.
Private local runOllama local + BGE/E5/nomic embeddings + FAISS/Qdrant localKeeps documents and inference local.
Best retrieval for support docsAdvanced RAG + Hybrid retrieval + Cohere/BGE rerankerBalances semantic meaning, exact terms, and final ranking.
Long document fidelityParent-Child RAG + document-aware chunking + MMRPreserves surrounding context while still retrieving precisely.
Strict release qualityAdvanced/Self-RAG + RAGAS + CI/CD thresholds + OpenTelemetryMeasures faithfulness, traces failures, and blocks regressions.

Simple Rules For Users

  1. Start with the simplest RAG that can answer the question.
  2. Fix extraction and chunking before changing the LLM.
  3. Use a new vector collection when changing embedding models.
  4. Add query enhancement when recall is poor.
  5. Add reranking when retrieved chunks include good and bad context together.
  6. Add compression when chunks are too long or noisy.
  7. Add evaluation before calling a system production-ready.

What To Avoid

Too much complexity early

Do not start with Agentic RAG, Self-RAG, reranking, compression, and fusion all at once. You will not know which part improved or broke quality.

Changing embeddings in the same index

This causes dimension mismatch or silent retrieval quality problems.

Skipping live smoke tests

Unit tests do not prove your Pinecone, Qdrant Cloud, MongoDB Atlas, or Azure setup works with real credentials.

Upgrade Path

  1. Start with a simple working baseline.
  2. Inspect retrieved chunks for failed queries.
  3. If relevant chunks are missing, improve extraction, chunking, query rewriting, or hybrid retrieval.
  4. If relevant chunks are present but low quality chunks are mixed in, add reranking.
  5. If chunks are too large or noisy, add context compression.
  6. If answers remain unsupported, use stricter prompts, Self-RAG, Corrective RAG, and faithfulness evaluation.
  7. Only after quality is measured, optimize latency and cost.
Next: Production readiness