AI RAG search with the Simple Data Analysis Library

This repository shows how to use the simple-data-analysis, maintained by Nael Shiab, to set up a free and open-source local RAG search, which allows you to search your data using natural language.

The default dataset is a list of AI-generated recipes. Here's an example query and result.

🔍 Question:
I am looking for a fruity pastry for breakfast.

📝 Answer:
The only pastry in the provided data that contains fruit and would be suitable for breakfast is **Apple Pie**. It uses sliced apples (such as Granny Smith or Honeycrisp) as the main fruit ingredient.

⏱️ Query duration: 8235ms
📊 Rows searched: 10
🧠 Thinking level: minimal
🤖 Model used: gpt-oss:20b

You can also return just the data, before it's passed to the LLM, which is way faster.

How it works

1. The data is loaded and converted to vectors (embeddings)
2. When a question is asked, two things happen in parallel:

• A vector search
  • The question is converted to vectors
  • The question vectors are compared the original data vectors
  • The closest data entries are kept
• A BM25 search
  • The words in the questions are searched in the original data text
  • This search takes into accounts how many times the words are found and how long are the data entries
  • The most relevant data entries are kept

3. Then the results of the vector and BM25 searches are reranked to prioritize search results that are appearing high with both search technics.
4. The selected data entries are sent to a LLM instructed to answer the question based only on the provided data, reducing the risk of hallucinations.

On a technical note, indexes are created for the embeddings and for the BM25 search. Data is cached locally. The reranking is a reciprocal rank fusion (RRF).

How to run

• Fork and then clone this repository

• Install Deno

• Install Ollama

• Pull the nomic-embed-text model for the embeddings: ollama pull nomic-embed-text

• Pull the gpt-oss:20b LLM model: ollama pull gpt-oss:20b

• Create an .env file from the .env.example

• Install all dependencies: deno install

• Run deno task data to load the data, create the embeddings and write the DB to disk.

• Run deno task server in one terminal to load and keep the DB in memory.

• Run deno task query in another terminal to ask a default question. The first time, it will load the LLM in memory. After that, it will be faster. Ollama keeps models in memory for 5 minutes by default.

Query Options

All query options can be combined. Here are the available parameters:

Basic query with custom question:

deno task query "I am looking for a pastry with herbal flavours."

Thinking level (-t or --thinking): Controls the depth of reasoning the LLM uses. Options: minimal, low, medium, high

deno task query "What dessert has chocolate?" -t high

Number of results (-n or --numDocs): How many semantically closest data points to pass to the LLM (default: 10)

deno task query "What dessert has chocolate?" -n 5

Endpoint (-e or --endpoint): Choose between query (default - uses LLM to answer) or data (returns just the search results without LLM, much faster)

# Get LLM-generated answer
deno task query "desserts with chocolate" -e query

# Get just the matching data without LLM processing
deno task query "desserts with chocolate" -e data

Combined example:

deno task query "I am looking for a pastry with herbal flavours." -t high -n 5 -e query

If you want to start fresh and remove the cache, run deno task clean.

Questions? Comments?

Reach out to Nael Shiab.