ResearchAssistant (RA)

RA is a local-first research assistant prototype built around an auditable project state machine. The core package owns project state, evidence, datasets, hypotheses, jobs, reports, and external-agent call records. Frontends are thin adapters.

Quick Start

conda activate RA
pip install -r requirements.txt
python -m adapter.cli.main ask "你的科研问题"
python -m adapter.cli.main local-ingest PRJ-xxxx
python -m adapter.cli.main local-search PRJ-xxxx "关键词"
python -m adapter.cli.main workflow "你的科研问题" --query "关键词"
streamlit run adapter/streamlit/app.py
uvicorn adapter.api.main:app --reload

Runtime project data is stored under projects/ by default.

The REST API is available through FastAPI at adapter.api.main:app. Start it with uvicorn adapter.api.main:app --reload, then use /health, /projects, and /projects/{project_id} endpoints for basic project operations.

Streamlit Workflow

• Chat: first message creates a project; later messages are recorded as auditable iterations with reports.
• Uploads: multiple files are classified and stored automatically. CSV/Excel/JSON files become profiled datasets; PDF/images/text/code/archive files become hashed artifacts.
• Experimental Data: confirm and process profiled datasets to create cleaned data, reproducible scripts, lineage hashes, and analysis evidence.
• Tools: run OpenAlex search, index/search uploaded project-local documents, generate initial hypotheses, and generate next-step plans.
• Tools: run the LangGraph MVP workflow for project input, local/online literature search, hypothesis generation, and next-step planning.
• API credentials are configured by file, not in the web UI. Copy mykey_template.py to mykey.py, then fill GenericAgent-style fields such as apikey, apibase, model, and api_mode. mykey.py and mykey.json are ignored by git.

MVP Scope

• Project creation and state transition.
• SQLite metadata and file-based artifacts.
• Evidence, hypothesis, iteration, job, and external-agent schemas.
• Structured external-agent output validation before any output is accepted as auditable project data.
• OpenAlex literature search with traceable paper/evidence records.
• Lightweight local RAG over uploaded project files and PDFs, with indexed chunks, page numbers where available, and evidence records.
• CSV/Excel upload profiling and reproducible dataset artifacts.
• Confirmed tabular processing with cleaned CSV outputs, processing scripts, lineage hashes, and analysis evidence.
• Basic numeric distribution and correlation plots for processed tabular datasets.
• Markdown iteration reports.
• CLI and Streamlit adapters.
• LangGraph workflow orchestration with SQLite/file-system business state preserved in research_agent_core.