Vector Databases - MS-RAGS(ALL-IN-ONE) Docs

Database Comparison

Vector databases are not interchangeable only by name. They differ in local vs managed operation, metadata filtering, hybrid search, index lifecycle, tenancy, backups, and the amount of infrastructure your team must own.

Database	Best for	Recommendation	Credentials
ChromaDB	Local development and demos	First choice for learning and local tests.	Optional `CHROMA_PERSIST_DIRECTORY`
FAISS	Fast local similarity search	Use for local file-persisted indexes.	Optional `FAISS_INDEX_PATH`
Pinecone	Managed production vector DB	Use when you want minimal operations.	`PINECONE_API_KEY`
Qdrant	Filtering, local/cloud flexibility	Strong production default.	`QDRANT_URL`, optional `QDRANT_API_KEY`
Weaviate	Hybrid search and schema-rich data	Use when object schema and hybrid search matter.	`WEAVIATE_URL`, optional `WEAVIATE_API_KEY`
Milvus	Very large vector collections	Use when scale is the main challenge.	`MILVUS_URI`, optional `MILVUS_TOKEN`
Redis	Low-latency vector search	Use when Redis is already part of the stack.	`REDIS_URL`
PGVector	PostgreSQL-native vector search	Use when relational consistency matters.	`PGVECTOR_CONNECTION_STRING`
Elasticsearch	Full-text plus vector hybrid search	Use for mature search teams.	`ELASTICSEARCH_URL`
OpenSearch	AWS/OpenSearch search stacks	Use for AWS-managed search environments.	`OPENSEARCH_URL`
Azure AI Search	Azure enterprise search	Use for Azure-based apps.	`AZURE_SEARCH_ENDPOINT`, `AZURE_SEARCH_KEY`
MongoDB Atlas Vector Search	MongoDB document applications	Use when source documents already live in MongoDB.	`MONGODB_ATLAS_CLUSTER_URI`

Production Rules

Match dimensions

The vector index dimension must match the selected embedding model. If the model changes, create a new collection or re-ingest.

Use names carefully

Collection names should identify environment, application, corpus, and embedding model family.

Smoke test live backends

Run opt-in smoke tests against the exact DBs planned for production.

How MS-RAGS(ALL-IN-ONE) uses Vector Stores

During setup

The CLI collects connection parameters, shows a summary, tests the connection, and asks for approval before ingestion begins.

During ingestion

Documents are loaded, split into chunks, embedded, and written into the selected collection or index with metadata attached.

During query

The retriever reads from the same populated vector store used during ingestion, avoiding the rebuild bug that can create empty runtime wrappers.

During generated-code use

Generated pipelines can ingest new documents or connect to an existing persisted collection for query-only usage when supported.

Collection Naming Strategy

Use names that tell future users what is inside the index.

msrag_{environment}_{app}_{corpus}_{embedding}

# Examples
msrag_dev_support_docs_openai_small
msrag_prod_policy_bge_m3
msrag_staging_legal_parent_child_3072

A good collection name reduces accidental ingestion into the wrong environment and helps users understand embedding compatibility.

Backend Selection Details

ChromaDB vs FAISS

Use Chroma when you want a simple local DB with collection semantics. Use FAISS when you want fast local vector search and file-backed persistence.

Pinecone vs Qdrant

Use Pinecone for managed serverless simplicity. Use Qdrant when filtering, local/cloud parity, and Rust-backed performance are attractive.

PGVector vs MongoDB Atlas

Use PGVector when your product data is relational. Use MongoDB Atlas when source documents already live in MongoDB collections.

Elasticsearch/OpenSearch/Azure AI Search

Use search platforms when hybrid lexical plus vector retrieval and enterprise search operations matter.

Smoke Test Examples

MS_RAG_SMOKE_VECTOR_DBS=chroma python -m pytest tests/smoke/test_vector_db_smoke.py -q
MS_RAG_SMOKE_VECTOR_DBS=faiss python -m pytest tests/smoke/test_vector_db_smoke.py -q
MS_RAG_SMOKE_VECTOR_DBS=qdrant python -m pytest tests/smoke/test_vector_db_smoke.py -q

Next: Recommendations