Ethiopian History RAG Assistant

A lightweight Retrieval-Augmented Generation (RAG) application that answers questions about Ethiopian history using a local text corpus, a persistent Chroma vector database, and a Groq-hosted LLM. It includes a Streamlit UI and an optional CLI.

Highlights

• Ingest .txt sources from data/, chunk and embed with sentence-transformers/all-MiniLM-L6-v2.
• Persist embeddings with Chroma (chroma_db/) for fast reloads.
• Retrieve top‑k relevant chunks via semantic similarity, with deduplication and a distance threshold.
• Build a concise, source-aware prompt for the LLM.
• Generate succinct answers via langchain_groq.ChatGroq authenticated by GROQ_API_KEY.
• Streamlit UI and CLI entry points.

Architecture

• code/app.py: Streamlit UI for querying and displaying answers.
• code/vectordb_and_ingestion.py: VectorDBManager (chunking, embedding, persistence, retrieval).
• code/prompt_builder.py: Builds compact prompts from retrieved chunks.
• code/retrieval_and_response.py: CLI pipeline (load/ingest → retrieve → prompt → answer).
• code/logger.py: Minimal console logger used across modules.
• wiki_fetcher.py: Fetches Wikipedia pages into data/ as .txt.

Prerequisites

• Python 3.12+
• Groq API key

Quickstart

# 1) Create and activate a virtual environment (recommended)
python3 -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate

# 2) Install dependencies
pip install -r requirements.txt

# 3) Configure environment
# Create .env and set GROQ_API_KEY
printf "GROQ_API_KEY=your_key_here\n" > .env

# 4) (Optional) Seed/expand the corpus with Wikipedia content
python -c "from wiki_fetcher import fetch_and_save_page; fetch_and_save_page('Axum Empire')"

# 5) Run the Streamlit app
streamlit run code/app.py

Screenshots

• Answer found from context:

• Answer not found (out-of-scope question):

CLI Usage

python -m code.retrieval_and_response "Who was Menelik II?" --top_k 2 --threshold 1.0

Environment Variables

• GROQ_API_KEY: Your Groq API key for langchain_groq.ChatGroq.

Example .env:

GROQ_API_KEY=your_key_here

Data & Persistence

• Place .txt sources in data/.
• On first run, the app/CLI chunks, embeds, and persists a Chroma collection in chroma_db/.
• Subsequent runs load the existing DB for fast startup.

End‑to‑End Flow

1. Load (or build) the Chroma vector database via VectorDBManager.
2. Perform a similarity search for the query to get candidate chunks.
3. Filter by distance threshold (lower = more similar) and remove duplicates.
4. Build a concise prompt via prompt_builder.build_prompt.
5. Generate an answer using langchain_groq.ChatGroq and display/return it.

Project Structure

rag-history-agent/
├─ code/
│  ├─ app.py                     # Streamlit UI
│  ├─ retrieval_and_response.py  # CLI entry
│  ├─ vectordb_and_ingestion.py  # VectorDBManager (chunk/embed/persist)
│  ├─ prompt_builder.py          # Prompt assembly
│  ├─ logger.py                  # Shared logger
│  └─ rag_assistant/             # Utilities (optional)
├─ wiki_fetcher.py               # Wikipedia ingestion utility
├─ data/                         # Text corpus (.txt files)
├─ chroma_db/                    # Chroma persistence (ignored by git)
├─ assets/screenshots/           # Images embedded in README/publication
├─ requirements.txt              # Python dependencies
├─ .gitignore
└─ README.md

Troubleshooting

• “GROQ_API_KEY missing”: Ensure .env exists, includes GROQ_API_KEY, and python-dotenv is installed.
• Long initial startup: First run embeds the corpus; subsequent runs are faster.
• Weak answers: Increase Top‑K, lower the distance threshold (e.g., 0.8–1.2), or add more documents to data/.
• Dependency errors: Verify the virtual environment is active and dependencies are installed from requirements.txt.

Suggested Dependencies

If requirements.txt isn’t populated, include:

streamlit
python-dotenv
langchain
langchain-community
langchain-groq
sentence-transformers
chromadb
requests
beautifulsoup4

Example Snippets

• Retrieval filter and deduplication:

# code/retrieval_and_response.py
results = vector_db.similarity_search_with_score(query, k=top_k * 5)
results = sorted(results, key=lambda x: (x[1] is None, x[1]))  # deterministic
unique_contexts, seen = [], set()
for doc, score in results:
    if score is not None and score > similarity_threshold:
        continue
    if doc.page_content not in seen:
        seen.add(doc.page_content)
        unique_contexts.append((doc, score))
    if len(unique_contexts) >= top_k:
        break

• Strict answer policy:

# code/prompt_builder.py
prompt = (
    "Answer ONLY using the context below.\n"
    "- If the context clearly contains the answer, state it concisely.\n"
    "- If the context does not contain the answer, reply exactly: I don't know\n\n"
    f"Context:\n{context}\n\n"
    f"Question: {query}\n"
    "Answer:"
)

• Deterministic LLM call (Groq):

from langchain_groq import ChatGroq

llm = ChatGroq(
    api_key=os.getenv("GROQ_API_KEY"),
    model="llama-3.1-8b-instant",
    temperature=0
)
response = llm.invoke(prompt)

Helpful Links

• Streamlit docs: https://streamlit.io/
• LangChain docs: https://python.langchain.com/
• Chroma: https://www.trychroma.com/
• Sentence Transformers: https://www.sbert.net/
• Groq: https://groq.com/

License

Provided as-is for educational purposes. For public release, consider the MIT License.

Contact me : aman.atalay-ug@aau.edu.et