Mnemosyne is an intelligent conversational search agent for Medium articles, named after the Greek goddess of memory.
Mnemosyne leverages Generative AI and other machine learning techniques to provide an intuitive, conversation-based interface for searching and exploring articles. Whether you're looking for specific information or wanting to dive deep into a topic, Mnemosyne acts as your personal search engine, offering relevant content and insights from a vast repository of your articles.
- Semantic Search through saved Medium Articles
- OpenAI and Ollama with Langchain integration for QnA
- Firecrawl integration for crawling articles
- MongoDB integration for building vector and text data store
- Dual mode operation with StreamingStdOutCallbackHandler and AsyncIteratorCallbackHandler()
- FastAPI and Quart support for flexible API deployment
- Streaming responses with SSE (Server-Sent Events)
-
Clone the repository:
git clone https://github.com/raghavpatnecha/mnemosyne.git cd mnemosyne -
Install dependencies:
You can install Conda for Python to resolve machine learning dependencies.pip install -r requirements.txt
The main configuration for the project is done in the src/config.py file. Here's how you can configure it:
-
MongoDB Settings:
Set up your MongoDB connection string and database name. -
OpenAI/Ollama Settings:
Provide your API keys for OpenAI and change the paramenters below based on your needs:class OPENAI: API_KEY: str = "" class LLM: MODEL_NAME: str = "gpt-4o-mini" #mistral,llama3.2 TOKEN_LIMIT: int = 125000 TEMPERATURE: float = 0.1 OPENAI_TIMEOUT: int = 20 ``` -
Firecrawl Settings:
Configure how often the system should crawl Medium for new articles.
You can run the application in Safe Mode or Unsafe Mode . Safe Mode ensures strict, reliable operations, while Unsafe Mode offers faster processing but with some risks.
FastAPI Server
Run the FastAPI server
python src/fast_api.py
Quart Server
Run the Quart server
python src/app.py
Search API:
GET /mnemosyne/api/v1/search/{query}?mode={sync} #Generates complete answer before streaming
GET /mnemosyne/api/v1/search/{query}?mode={async} #Uses AsyncStreamingCallbackHandler , lower Latnecy
- config.py: Configuration file containing settings for MongoDB, API keys, and other integrations.
- app.py/fast_api.py: app.py/fast_api.py: Web server implementations using Quart and FastAPI respectively
- LLMService.py: Uses factory and Strategy design patterns to ensure extensibility and maintainability:, contains code for text stream generation using langchain.
- MnemsoyneService.py: Manages the overall Mnemosyne service, coordinating between search.py and llmservice.py.
- script.js = Uses highlight.js and marked.js to parse repsone on frontend
mnemosyne/
├── src/
│ ├── app.py # Quart server implementation
│ ├── config.py # Configuration settings
│ ├── fast_api.py # FastAPI server implementation
│ ├── api/
│ │ └── search.py # Core search API implementation
│ ├── model/
│ │ └── model_utils.py # Utility functions for models
│ ├── service/
│ │ ├── LLMService.py # LLM integration service
│ │ ├── llm_utils.py # LLM utility functions
│ │ ├── MnemsoyneService.py # Main service coordination
│ │ ├── MongoService.py # MongoDB service
│ │ └── mongo_utils.py # MongoDB utility functions
│ ├── static/ # Static files
│ └── templates/ # HTML templates
This project is licensed under the MIT License - see the LICENSE.md file for details.
For any queries or suggestions, please open an issue on this GitHub repository or contact the maintainers directly.
@article{raghavpatnecha_mnemosyne,
author = [{Patnecha, Raghav}, {Bahadur, Akshay}],
journal = {https://github.com/raghavpatnecha/Mnemosyne},
month = {10},
title = {{Mnemosyne}},
year = {2024}
}