Skip to main content
Latest research:When the Circuit Dissolves →12 vindexes on Hugging Face
Request demo

RAG Arena &
Dynamic Routing

Compare multiple knowledge bases side-by-side, discover which performs best for each question type, and let the system automatically route future queries to the winning source.

Side-by-Side Knowledge Comparison

Send a question to multiple RAG configurations simultaneously. See how different knowledge bases respond, then pick the winner.

"What are the recommended dosing guidelines for this medication?"
Variant A QdrantQdrant
Retrieving context...
Based on the prescribing information (rev. 2024-Q3), the recommended starting dose is 10 mg once daily, with or without food. Titrate up to a maximum of 40 mg/day in 10 mg increments at minimum 2-week intervals based on therapeutic response.

Monitoring during the first 8 weeks: serum creatinine + eGFR, hepatic enzymes (ALT/AST), and blood pressure — at baseline, week 4, week 8, then quarterly thereafter.

Dose adjustments: No adjustment for mild renal impairment (eGFR > 60). Moderate-to-severe (eGFR 30–59) requires a 50% reduction. Severe hepatic impairment (Child-Pugh C): not recommended.

Interactions: Strong CYP3A4 inhibitors are contraindicated. With moderate inhibitors, limit to 20 mg/day.

Sources: clinical-guidelines-2024.pdf p.142–146 · pharmacology-handbook.pdf p.78
Variant B CloudflareCloudflare Vectorize
Retrieving context...
Dosing should begin at the lowest effective dose. The product label indicates 10–20 mg as the typical adult range. Patients should be monitored for adverse effects during the first two weeks of therapy.

Common side effects: nausea, headache, and mild dizziness — typically transient and resolve within the first month. Severe reactions are rare (<0.1%) but include hepatotoxicity; discontinue immediately if jaundice or unexplained transaminase elevation occurs.

Discontinuation: Taper gradually over 1–2 weeks to avoid withdrawal symptoms. The drug is metabolized primarily by CYP2D6; poor metabolizers may require lower doses or alternative agents.

Sources: prescribing-info-v3.pdf p.12 · safety-monitoring.pdf p.34–37

How RAG Arena Works

A structured process for finding the best knowledge base configuration for every question type.

1

Configure Variants

Set up 2-8 arena variants, each pointing to a different RAG vector or knowledge base.

2

Run in Parallel

Send questions to all variants simultaneously. Each retrieves context from its own knowledge base.

3

Compare & Score

View responses side-by-side with inline quality scores for relevance, hallucination, and completeness.

4

Learn & Route

The system learns which knowledge base wins for each question type and routes future queries automatically.

Dynamic RAG Routing

Once arena results establish preferences, incoming questions are automatically routed to the best-performing knowledge base — no manual configuration needed.

Incoming
User Question
Semantic Match
Route Lookup
Best Match
Qdrant (0.94)
Response
Optimized Answer

Built for Enterprise RAG

Everything you need to optimize retrieval quality at scale.

Multi-Provider Support

Compare across Qdrant, Cloudflare Vectorize, Couchbase, MongoDB Atlas, and more in a single arena experiment.

Inline Quality Scoring

Every response is scored for relevance, hallucination, correctness, and completeness in real time.

Auto-Fix Integration

Arena results feed into the QA auto-fix loop, automatically learning optimal routing with quality thresholds.

Full Audit Trail

Every routing decision is tracked with source (arena, auto-fix, or manual), scores, and timestamps in MongoDB.

Fast Routing via KV

Learned preferences stored in Cloudflare KV for sub-millisecond lookups. MongoDB maintains the full history.

Arena Presets

Save, name, and reuse arena configurations. Export as portable JSON for sharing across teams.

Supported Vector Providers

Run arena experiments across any combination of vector databases.

QdrantQdrant
CloudflareCloudflare Vectorize
CouchbaseCouchbase
GoogleVertex AI Vector Search
MongoDB Atlas
Redis
PageIndexPageIndex

RAG Retrieval Architectures

Route queries to different retrieval strategies based on question type. Each architecture optimizes for different use cases.

VECTOR SEARCHQEmbedR

Vector Search

Semantic similarity matching via dense embeddings. Best for natural language questions and conceptual lookups.

HYBRID SEARCHQVectorBM25RRF

Hybrid (BM25 + Vector)

Combines keyword matching with semantic search using Reciprocal Rank Fusion. Best for precise technical queries.

RERANKINGQRetrieveCross-encoderrerank3

Reranking

Initial retrieval followed by cross-encoder reranking for precision. Best for high-stakes answers requiring accuracy.

AGENTIC RAGQAgentRouterKB 1KB 2ToolsSyn

Agentic RAG

LLM-driven routing across multiple knowledge bases and tools. Best for complex multi-step questions requiring synthesis.

Find your optimal RAG configuration

Stop guessing which knowledge base performs best. Let RAG Arena show you the data, then let Dynamic Routing handle the rest.