• 

• 

• 

• 

• 

• 

Cosmic Engine

Cosmic Engine is a RAG experimentation platform for comparing 10+ retrieval, chunking, generation, and caching strategies.

Most RAG tools hide retrieval decisions.

Cosmic Engine exposes retrieval, ranking, caching, and generation behavior so you can improve quality with measurable evidence.

All those fine-tuning strategies you found lying around the internet, thanks to the Anthropic blogs, are put into a single place to experiment with.

Retrieval Modes

Cosmic Engine comes packed with retrieval modes that can be toggled per session:

• Vector Search: Dense embedding retrieval from Milvus.
• Hybrid Search: BM25 + vector retrieval with rank fusion.
• RRF Search: Reciprocal Rank Fusion over lexical and dense ranked lists.
• Query Expansion: Optional LLM-assisted expansion of vague or short queries before retrieval.
• Query Response Caching: Normalized cache keys based on session, query, attachment set, and retrieval options. Repeated queries can return cached assistant responses quickly.
• Dynamic Chunk Retrieval Sizing: Top-K automatically adapts to the number of selected attachments. This one's my favourite.
• And many more.

Retrieval Diagnostics

Per-attachment retrieval timing and diagnostics record:

• Vector search timing
• BM25 timing
• Ranking timing
• Document hydration timing

This makes it easier to understand where your retrieval pipeline is spending time and how different strategies affect performance.

Business Value Drivers

• Faster answer discovery from large, unstructured document sets.
• Lower hallucination risk through retrieval-grounded responses and citation workflows.
• Better quality tuning through measurable retrieval and generation metrics.
• Reduced repeated inference costs when caching is enabled.

Production Hardening

For larger commercial rollouts, you would typically want to add:

• Centralized or distributed queue backend
• Rate limiting and abuse controls
• Strong observability and alerting
• Multi-tenant isolation strategy
• SSO and audit controls

Tech Stack

Frontend

• Next.js
• React
• TypeScript

Backend API

• Express 5
• TypeScript

Ingestion Microservice

• FastAPI
• Python
• MarkItDown

LLMs and Embeddings

• Google GenAI APIs
• Bring Your Own API Key

Data

• Milvus for vector storage
• PostgreSQL + Prisma for transactional data

Open Source

Cosmic Engine is open source.

Check it out and experiment with the different retrieval strategies:

github.com/arpan-lol/cosmic-engine

Built With

• genai
• milvus
• typescript