End-to-End: Document Q&A Mini Project

The RAG Pipeline

Retrieval-Augmented Generation (RAG) answers questions using your own documents: chunk, embed, store, retrieve, and let the model answer from the retrieved text.

Step 1 - Prepare Chunks

documents = [
    "GroovyGrails teaches Java from zero to advanced.",
    "Spring Boot makes Java web apps fast to build.",
    "Groovy is a JVM language that pairs with Java.",
    "Vector databases power retrieval in AI assistants.",
]
chunks = [documents[i:i] for i in range(len(documents))]

Step 2 - Embed and Store

from sentence_transformers import SentenceTransformer
import chromadb

model = SentenceTransformer("all-MiniLM-L6-v2")
client = chromadb.PersistentClient(path="kb")
kb = client.get_or_create_collection("kb", metadata={"hnsw:space": "cosine"})

kb.add(
    ids=[f"c{i}" for i in range(len(chunks))],
    documents=chunks,
)

Step 3 - Retrieve

query = "Which framework builds Java web apps quickly?"
hits = kb.query(query_texts=[query], n_results=2)
context = " | ".join(hits["documents"][0])
print(context)

Step 4 - Generate the Answer

prompt = f"Answer using only the notes below.

{context}

Question: {query}"
# result = call_your_llm(prompt)   # plug in your model or API key
# print(result)

The model receives only the retrieved notes, so answers stay grounded in your content.

Improving the Project

• Use overlapping chunks of a few hundred characters.
• Add metadata such as source titles and dates.
• Return more chunks and rerank them before generation.
• Ask the model to cite the chunk it based each part on.

Key Points

• RAG = chunk, embed, retrieve, generate.
• Retrieved context keeps answers grounded.
• Persist the knowledge base so you build once and reuse.
• Reranking and chunk tuning raise answer quality.