Try it out!
https://hindict.anuragshenoy.in/
Demo
I have used a local Large Language Model (hereafter referred to as an LLM) to provide the user with the basic translation.
Then, I'm extracting the word guessed by the LLM to call a free dictionary API to get a full definition of the English term.
Backstory
To learn more about civilizations and hopefully more about myself, I read Genesis (from the Bible) a long time ago. I was always interested in the Bhagavad Gita as a source of wisdom and philosophy, but never picked it up.
Watching Oppenheimer last year pushed me to finally take the plunge and start reading it. The quote everyone knows by now, piqued my interest.
I purchased a Hindi version of the Bhagavad Gita to be closer to the original Sanskrit version and not be heavily influenced by the English language when it came to distilling ideas difficult to accurately translate and condense.
This posed a new challenge: My Hindi isn't the greatest, and I'm much slower at reading and writing Hindi compared to speaking the language.
I have always used the latin alphabet to write Hindi when chatting with friends.
To make it much easier to learn Hindi words quickly, I decided to make an app which uses the power of generative AI to provide me the meanings of words directly from the romanized version of the Hindi word.
I then use an API to fetch the definition of the top English word which is synonymous to the Hindi word.
Future plans
The plan is to extend this to have a full Hindi dictionary entry of the word searched in addition to the English definition of the closest equivalent.
Right now, I couldn't find a Hindi dictionary with a permissive license.
I'm considering parsing Wiktionary's data dump to generate my own Hindi dictionary.
Prompt Engineering
Respond in one concise sentence without disclaimers or examples. Format: User word title-case, (user word in Hindi), refers to … (provide a definition in 5-15 words). Always conclude with this exact phrase: 'Closest English equivalent: (closest english equivalent single word)'. Do not omit this phrase under any circumstances.
Had to add the "Do not omit this phrase under any circumstances" because the model was hallucinating.
This allows me parse the closest English equivalent word and retrieve it's definition.
More demo images
Bhumi
Bhumi (dark mode)
Gyan
Manushya
Janma (dark mode)
Tools used
Ollama
Ollama can be used to serve LLMs locally. I'm using the Ollama docker container to containerize the model and simplify deployment.
You could run it directly on your machine to achieve a small performance improvement over the docker container.
Llama 3.2 3B token model
https://ollama.com/library/llama3.2:3b
https://www.llama.com/llama-downloads/
You'll need permission to use the models.
Open AI 4o-mini
Using OpenAI's GPT 4o-mini in the production version of the app.
ReactJS + NextJS
Next JS for the frontend to make a quality single-page-application (SPA).
References and Further Reading
- Course: Foundations of Local Large Language models by Duke University https://www.coursera.org/learn/local-large-language-models/
- Ollama https://github.com/ollama/ollama?tab=readme-ov-file#ollama
- Free Dictionary API https://github.com/meetDeveloper/freeDictionaryAPI
- Indic transliteration tools https://github.com/indic-transliteration/indic_transliteration_py?tab=readme-ov-file#indic-transliteration-tools
- FreeDict https://freedict.org/downloads/#dictionary-downloads
