elimu-research-assistant 1.2.5

An intelligent research assistant for Kenyan educators to create localized, contextual educational content that bridges the context deficit in education.

Elimu Research Assistant

An AI research agent for Kenyan educators that generates locally-grounded educational content using a ReAct (Reasoning + Acting) loop over live web search.

What It Does

Given a teaching request such as "Create a Form 3 Business Studies lesson on M-Pesa's impact on small enterprises", the agent:

1. Reasons about what to search and in what order
2. Executes web searches via Serper.dev, biased toward .ke and East African domains
3. Scrapes and cleans source pages for relevant content
4. Iterates (up to 12 reasoning steps by default) until it can synthesize a cited, structured educational artefact
5. Returns classroom-ready output: lesson plan, case study, handout, or assessment, depending on the request

Architecture

CLI / Streamlit
      │
      ▼
build_elimu_agent()          ← elimu_react/__init__.py
      │
      ├── LLMInterface        ← Google Gemini 2.x (primary: gemini-2.0-flash)
      │     └── model fallback chain (2.0-flash → 2.5-flash → 2.5-pro)
      │
      ├── ToolManager
      │     ├── SearchTool    ← Serper.dev (Google Search API)
      │     │     └── Kenya bias: appends "Kenya", prioritises .ke / .ac.ke
      │     └── ScrapeTool    ← requests + BeautifulSoup, 6 000 char limit
      │
      └── ReActAgent          ← elimu_react/agent.py
            └── run(task) → ReAct loop (max 15 iters)
                  Thought → Action → Observation → … → Final Answer

ReAct Loop (elimu_react/agent.py)

Each iteration:

1. Build a prompt: system instructions + tool descriptions + task + prior steps
2. Call LLM → parse Thought, Action, Action Input, or Final Answer
3. If action: dispatch to SearchTool or ScrapeTool, collect observation
4. Append step to trace; repeat
5. On Final Answer or exhausted iterations → return synthesis

The system prompt instructs the model to produce classroom-ready artefacts with Kenyan examples and cited sources, and to never fabricate URLs or statistics.

Configuration (config/config.py)

Credentials and settings are resolved in priority order: .env → environment variables → system keyring. Non-sensitive settings persist in ~/.elimu_research_assistant/config.json.

Key	Default	Description
gemini_api_key	—	Required. Google Gemini API key
serper_api_key	—	Required. Serper.dev API key
model_name	gemini-2.0-flash	Primary LLM model
model_fallback	gemini-2.5-flash	Fallback on 404/quota
model_temperature	0.15	Low temperature for factual output
max_iterations	12	ReAct loop cap
max_tool_output_length	6000	Observation truncation (chars)
educational_focus	True	Adds "classroom" to education queries
prioritize_kenyan_sources	True	Biases search toward .ke domains

Installation

git clone https://github.com/ashioyajotham/elimu_research_assistant.git
cd elimu_research_assistant
pip install -e .

API Keys

You need two keys:

• Google Gemini: aistudio.google.com — free tier available
• Serper.dev: serper.dev — 2 500 free searches/month

elimu config --api-key YOUR_GEMINI_KEY
elimu config --serper-key YOUR_SERPER_KEY

Keys are stored in the Windows system keyring by default, or in .env if keyring is unavailable.

CLI Usage

# Single research task
elimu research "Create a Form 3 lesson on M-Pesa's impact on small enterprises"

# Batch from file (one task per line or blank-line-separated blocks)
elimu batch-research tasks.txt --output results/

# Interactive shell with history and autocomplete
elimu shell

# Configuration
elimu config --show
elimu config --model gemini-2.5-pro
elimu config --format markdown   # markdown | json | html

Output files are saved to results/result_<sanitized-query>.md by default.

Streamlit Webapp

pip install streamlit
streamlit run streamlit_app.py

The webapp exposes the same research agent with a browser UI: sidebar configuration, research input, expandable ReAct trace, and inline result rendering.

Project Structure

elimu_research_assistant/
├── elimu_react/          # ReAct agent engine
│   ├── agent.py          # ReActAgent: reasoning loop
│   ├── llm.py            # LLMInterface: Gemini wrapper with fallback
│   └── tools/
│       ├── search.py     # SearchTool (Serper.dev)
│       └── scrape.py     # ScrapeTool (requests + BS4)
├── config/
│   ├── config.py         # ElimuConfigManager: env / keyring / file
│   └── config_manager.py # BaseConfigManager
├── utils/
│   ├── console_ui.py     # Rich theme, panels, progress
│   ├── react_output.py   # Markdown serialiser for ReAct traces
│   └── task_parser.py    # Multi-line task file parser
├── cli.py                # Click CLI entry point
├── streamlit_app.py      # Streamlit webapp
└── pyproject.toml

Development

pip install -r requirements.txt
python -m pytest

Logs are written to logs/agent.log. Set --verbose on any CLI command for INFO-level output.

License

MIT — see LICENSE.

Author

Ashioya Jotham · [e-mail]