A Shrialok AI Labs product
GroundGate: Production RAG Kit for Python
RAG that says "I don't know" instead of guessing.
Free Lite version, plus paid Starter, Pro and Agency versions
Get GroundGate on GumroadWhat is GroundGate?
GroundGate is a production-grade retrieval-augmented generation (RAG) backend in Python built around one rule: answer only from your documents, cite every claim, and refuse when the answer is not there.
Most RAG starter kits are chat-with-a-PDF demos. They work in the demo and then, in production, give confident wrong answers, miss exact error codes, and quietly get worse every time someone changes a chunk size. GroundGate is built to catch those failures.
What it does
- Hybrid retrieval: vector and BM25 keyword search fused with Reciprocal Rank Fusion, so exact identifiers such as error codes, SKUs and versions are not lost
- Reranking: a dependency-free reranker built in, with a local cross-encoder, Cohere or an LLM reranker available through one setting
- Refuses instead of inventing: a relevance gate stops the LLM from answering when no document is relevant
- Claim-by-claim grounding: every sentence is checked against the source it cites, and unsupported claims are removed
- Prompt-injection defence: hidden HTML is dropped, documents with injected instructions are quarantined, and a canary token catches prompt leaks
- Multi-tenant by design: tenant filtering before search, plus Postgres row-level security SQL
- Runs with OpenAI or any OpenAI-compatible API, Anthropic Claude, fully local Ollama, or offline with no API key
- Production plumbing: FastAPI API, web console, Postgres with pgvector and Docker Compose, caching, and per-request traces with tokens and cost
Who it is for
- Freelancers and agencies building RAG for clients
- Backend developers moving into AI engineering
- Indie hackers who need document Q&A they can trust
Frequently asked questions
How do I stop a RAG system from hallucinating?
GroundGate uses two checks. A relevance gate refuses to answer when no retrieved document is relevant, and a grounding step checks every sentence of the answer against the source it cites and removes unsupported claims.
Do I need an API key?
No. It runs offline for testing and fully locally with Ollama. It also works with OpenAI, any OpenAI-compatible API, and Anthropic Claude.
Is GroundGate a hosted service?
No. You get the Python source code and run it on your own machine or server.
Does it use LangChain or LlamaIndex?
No. It is plain Python with FastAPI, httpx and numpy, so every line is readable and changeable.
Is there a free version?
Yes. GroundGate Lite is free and open source under the MIT license. The paid Starter, Pro and Agency versions add the full pipeline, the evaluation suite and a CI quality gate.
Can I use it for client projects?
Starter and Pro cover your own products and projects you build yourself. Agency covers your whole team and delivering the code to clients. Reselling the kit is not allowed.
What do I need to run it?
Python 3.10 or newer. Docker, an OpenAI or Anthropic API key, and Ollama are optional.
RAG that says "I don't know" instead of guessing.
Free Lite version, plus paid Starter, Pro and Agency versions. Checkout and delivery are handled by Gumroad.
Get GroundGate on Gumroad