Researcher: Agentic Writing Assistant

Researcher is an autonomous agent powered by Mistral LLMs that helps you iteratively write and improve scientific manuscripts. It uses a set of integrated tools (retrieval, critique, file I/O, thinking) and an interactive loop to generate, revise, and manage your manuscript based on your data.

Key Features

• 🤖 Autonomous Agentic Workflow: The agent generates content, solicits self-critique, and applies feedback.
• 🧠 Mistral LLM Integration: Chat and embedding models for content generation and retrieval.
• 📚 Retrieval Augmented Generation: Contextual vector search over your document corpus.
• 🔧 Tool-System: list_files, retrieve_relevant_documents, critique_content, think, read_from_file, write_to_file.
• 💾 Persistent State with Weave: Save and resume sessions, maintain conversation history, and manage manuscript file.
• 📄 Manuscript Management: Automatic handling of workdir/manuscript.txt, backups, and versioning.
• 🚀 Document Preprocessing: Chunk PDF documents into JSONL for indexing via researcher.prepare.
• 🔄 Interactive REPL: Command-line interface to interact with the agent, provide initial prompts, and exit gracefully.

Installation

git clone https://github.com/tcapelle/simple_agent.git
cd researcher
pip install -e .

Make sure to set your Mistral API key:

export MISTRAL_API_KEY=your_api_key_here

Usage

Preprocess Documents and Build Vector DB

Process all PDFs in your data directory and build the contextual vector database:

researcher.prepare

Start the Agent

Use the researcher entrypoint to launch the interactive manuscript assistant:

researcher

• On first run, if no database is found, you can run researcher.prepare (as above) to generate the preprocessed documents and vector DB.
• The agent will prompt you for an initial manuscript prompt or resume an existing one. Use exit or quit to end the session.

License

MIT