rattus

Conversational AI for sales insights.

Getting started

1. Open terminal.
2. Navigate to the root directory of the application.
3. Setup virtual environment using python -m venv env
4. To activate env run env\Scripts\activate or source env/bin/activate.
5. Run pip install -r requirements
6. Update config.py file:

O4_MINI_DEPLOYMENT_NAME = "o4-mini"
O4_MINI_API_KEY = <<llm-api-key>>
AZURE_OPENAI_ENDPOINT = <<llm-model-url>>

TEXT_EMBED_API_KEY = <<embedding-model-key>>


DIMENSIONS = 3072
COSMOS_URL = <<cosmos-db-url>>
COSMOS_KEY = <<cosmos-db-api-key>>

CONTEXT_DB = <<context-cosmos-db-name>>
SCHEMA_CONTAINER = <<schema-cosmos-container-name>>
DATA_CONTAINER = <<data-cosmos-container-name>>


AZURE_OPENAI_MODEL=O4_MINI_DEPLOYMENT_NAME

LANGSMITH_API_KEY = <<your-langsmith-api-key>>

6. Create a new ipynb file in root directory, preferably with name checks.ipynb.
7. Set kernel to local environment.
8. Create a new code cell and run the following code:

%load_ext autoreload
%autoreload 2

from ai.graph import run_graph

response = run_graph(user_query="Which is the most sold product?")
response

pip 25.3 (python 3.14)

Assumptions

• The data/reports would be contain columnar data.
• columns in sheet would not be too high to break the schema vector.
• reports dont have a column with name "vector"
• The data type of column could be determined by the first 5 rows.

Limitations & Unaddressed:

• Cosmos DB is choosen as primary data storage, it fails for large datasets. Cosmos DB container, each logical partition is capped at 20 GB of storage, and each physical partition (where data lives) scales up to 50 GB as needed, automatically handled by Cosmos DB based on your throughput and data size.
• Language and unicodes are ignored.
• Location handled by vectors.
• Type of data under each column isn't strictly maintained or tracked.
• Cosmos DB failing at several arthematic operations, proving unreliable. Need for another big data suitable DB.
• Date system. Fails to identify YY-MM-DD vs YY-DD-MM.
• No infinite loop handling implemented.
• No guardrails
• Messages in chat are not cached.