trace
$ app.py ← active Streamlit session
$ read query + limits from config
01
The RAG loop: query to grounded answer
This path runs in real time on every user question: embed, search, filter, answer. Walk through the steps below.
Progress
Step 1 / 6
02
Architecture explorer
Click a line under src/ or a tile to see what that section of the codebase is for.
Repository layout
rag-starter-kit/
├── src/
├── data/ # your Markdown corpus
└── .env.example
Modules (click a tile or a line under src/)
03
End-to-end ingestion path
Batch RAG ingestion pipeline: I run mine on a schedule (e.g. weekly) so new posts are ingested into the database. Tap a stage for details.
I read data from markdown files and chunk them.
Transform chunked content into embeddings.
Write text and vectors into the database.
Once you have data, the app is ready to retrieve and answer questions. No separate training or fine-tuning needed.
04
Configurations
Below are configurations you can adjust to match your content, models, and budget. Tap a key for what each setting does.
Models
Choose your embedding model.
Configure model choice, randomness, max length: cost, latency, truncation.
Retrieval
How close context must match query.
How many relevant sources to include as context for LLM system prompt.
Ingest
Size of each chunk when ingested into vector database.
Optional summaries for source previews.
05
Ship your own AI assistant grounded in your content
Plug-and-play stack
Follow the guide and deploy an AI assistant grounded in your content in ~60 minutes with these steps:
From clone to deploy
1
Clone & configure
2
Set up database
3
Set up credentials
4
Ingest your data
5
Test system
6
Deploy to live URL
Get the Implementation Guide on Substack (paid tier)
I've documented the exact steps that you can use to deploy an assistant on your own content.
Deploy your own grounded assistant in ~60 minutes when you follow the guide.
Follow the implementation guide