RAG Providers

Detailed configuration for each supported vector database provider.

---

Azure AI Search

Microsoft's managed search service with native semantic search, reranking, and hybrid search capabilities.

Best for: Azure-native deployments, enterprise environments with Azure infrastructure, use cases needing hybrid (keyword + semantic) search.

Required Fields

Field	Example	Notes
Endpoint	https://my-search.search.windows.net	From Azure Portal → AI Search service → Overview
API Key	abc123...	Admin or Query key from Azure Portal → Keys
Index Name	product-docs-index	The name of your search index
Semantic Config	my-semantic-config	(Optional) For semantic ranking
Vector Field	contentVector	Field in your index containing vectors
Content Field	content	Field containing the text to retrieve

Setup Notes

• Create an Azure AI Search resource in Azure Portal
• Create an index with a vector field (1536 dimensions, cosine similarity, HNSW algorithm)
• Index your documents using the Azure SDK or REST API
• Get the endpoint from the Overview blade
• Get the API key from the Keys blade (use Query key for read-only access)

Free Tier

Azure AI Search has a free tier (F0) with 50MB of storage — fine for testing. For production, use Basic (2GB) or Standard (25GB+).

Hybrid Search

If your index has both keyword and vector fields, Azure AI Search supports hybrid search (combines keyword + semantic). Configure this via the API if needed.

---

Pinecone

Popular managed vector database. Known for its simple API and serverless tier.

Best for: Getting started quickly, production-scale deployments, serverless architecture.

Required Fields

Field	Example	Notes
API Key	pc-abc123...	From Pinecone Console → API Keys
Index Name	product-docs	The index to query
Host	product-docs-xxxxx.svc.pinecone.io	From Pinecone Console → Index → Host
Namespace	default	(Optional) Namespace within the index

Setup Notes

• Sign up at pinecone.io
• Create an index with 1536 dimensions and cosine metric
• Index name must match exactly (case-sensitive)
• For serverless indexes, get the host from the index details page
• For starter tier, note the request rate limits

Free Tier

Pinecone's starter plan includes one index, 100MB storage, and limited requests — good for development. Production typically needs Standard plan.

Namespaces

If you're using namespaces to partition your data (e.g., one namespace per client), enter the specific namespace. Leave empty to query the default namespace.

---

Weaviate

Open-source vector database with strong filtering and multi-modal support.

Best for: Self-hosted deployments, complex filtering (where queries), multi-tenant architectures.

Required Fields (Cloud)

Field	Example	Notes
URL	https://my-instance.weaviate.network	From Weaviate Cloud Console
API Key	weaviate-abc123	From Weaviate Cloud Console → Details
Class Name	ProductDocument	Weaviate class to query (note: PascalCase)
Text Field	content	The property containing the text
Vector Field	—	Weaviate computes vectors via its modules

Required Fields (Self-hosted)

Field	Example	Notes
URL	https://weaviate.yourcompany.com	Your Weaviate server URL
API Key	—	(Optional, if auth enabled)
Class Name	ProductDocument
Module	text2vec-openai or text2vec-transformers	Embedding module configured on the server

Setup Notes

• Weaviate classes are like tables — create a class for your content
• The class schema defines properties and which module handles embedding
• For OpenAI embedding module (text2vec-openai), Weaviate handles embedding — you need OpenAI config in Weaviate
• For TARX's external embedding approach, configure Weaviate to accept pre-computed vectors

---

Qdrant

High-performance open-source vector database with Rust core.

Best for: High-throughput production workloads, self-hosted deployments, complex filtering.

Required Fields

Field	Example	Notes
URL	https://my-cluster.cloud.qdrant.io	Qdrant Cloud URL or self-hosted endpoint
API Key	qdrant-abc123	From Qdrant Cloud Console, or set in self-hosted config
Collection	product_docs	The Qdrant collection to query
Vector Name	content	Name of the named vector (if using named vectors)

Setup Notes

• Create a collection with 1536 dimensions and cosine distance
• When uploading points, include the text content in payload: {"content": "your text", "source_url": "https://..."}
• Qdrant Cloud has a free tier for development
• For self-hosted, ensure the API is accessible from TARX's backend

Payload Filtering

Qdrant supports rich payload filtering. If your collection has metadata (e.g., category, date), you can add filters. Currently supported via the Custom REST provider type for advanced filtering.

---

Supabase Vector

PostgreSQL-based vector storage using the pgvector extension.

Best for: Teams already using Supabase, SQL-native vector operations, combining vector search with relational data.

Required Fields

Field	Example	Notes
Supabase URL	https://abcxyz.supabase.co	From Supabase Dashboard → Settings → API
Service Role Key	eyJhbGc...	Service role key (NOT anon key)
Table Name	documents	PostgreSQL table with vector column
Content Column	content	Column containing the text
Embedding Column	embedding	Column with vector(1536) type
Metadata Columns	source_url,title	Comma-separated columns to return as metadata

Setup Notes

• Enable pgvector extension in Supabase (Database → Extensions → vector)
• Create a table:

  CREATE TABLE documents (
    id bigserial primary key,
    content text,
    embedding vector(1536),
    source_url text,
    title text,
    created_at timestamp default now()
  );
  
  CREATE INDEX ON documents USING ivfflat (embedding vector_cosine_ops)
    WITH (lists = 100);

• Use the service role key for server-side queries (has full access)
• TARX performs: SELECT content, source_url, 1 - (embedding <=> $1) as score FROM documents ORDER BY score DESC LIMIT $2

---

Custom REST

For any vector database that exposes a REST API. This is the escape hatch for providers not natively supported.

Best for: Custom vector database deployments, internal enterprise vector stores, niche providers.

Required Fields

Field	Example	Notes
Search URL	https://my-vectordb.com/api/search	Endpoint that accepts search queries
Method	POST	HTTP method (usually POST)
Auth Header	Authorization: Bearer {api_key}	How to pass credentials
Request Template	See below	JSON template for the search request
Response Path	$.results	JSONPath to extract results from response
Text Path	$.results[*].text	JSONPath to the text content in each result
Score Path	$.results[*].score	JSONPath to similarity scores

Request Template

TARX sends the embedded query vector as {vector} in the request template:

{
  "vector": "{vector}",
  "top_k": "{top_k}",
  "namespace": "my-namespace"
}

TARX substitutes {vector} with the actual embedding array and {top_k} with the configured value.

Response Parsing

Your endpoint must return JSON. TARX uses JSONPath to extract:

• The text content of each result
• The similarity score

Example expected response:

{
  "results": [
    {"text": "SSO setup requires...", "score": 0.94, "url": "https://docs.example.com/sso"},
    {"text": "SAML configuration...", "score": 0.87, "url": "https://docs.example.com/saml"}
  ]
}

---

Provider Comparison

	Azure AI Search	Pinecone	Weaviate	Qdrant	Supabase	Custom REST
Managed cloud	✅	✅	✅	✅	✅	Varies
Self-hosted	❌	❌	✅	✅	✅	✅
Free tier	✅	✅	✅	✅	✅	Varies
Hybrid search	✅	❌	✅	✅	❌	Varies
Native Azure	✅	❌	❌	❌	❌	❌
SQL familiarity	❌	❌	❌	❌	✅	Varies
Setup complexity	Medium	Low	Medium	Medium	Medium	High