A Hinglish evil-twin chatbot
Streamlit chatbot with a roast-style Hinglish persona, local Ollama or OpenAI-compatible LLMs, and document Q&A with source labels.
KaleshAI is a mischievous Hinglish chatbot with a real LLM-backed chat flow and optional document Q&A. It is an evil-ish twin: quick roasts, light gaali, and it still answers the question.
Runs on a local Ollama model by default, or any OpenAI-compatible endpoint you configure.
Upload TXT, MD or text-based PDF; text is chunked and retrieved with lexical BM25, then passed to the LLM.
Answers from documents show which source chunks they came from.
Provider failures (HTTP errors, timeouts, bad replies) raise clear errors instead of crashing the chat.
20 unit tests with a mocked LLM client, so no paid model calls.
Local Ollama, or an OpenAI-compatible base URL, model and key.
Upload files; they are chunked and indexed with BM25.
Matching chunks go to the LLM with the roast persona; answers carry source labels.