Part 14 — RAG: Give Your AI a Memory

def chunk_text(text, chunk_size=500, overlap=50):
    words = text.split()
    return [' '.join(words[i:i+chunk_size])
            for i in range(0, len(words), chunk_size - overlap)]
# Chunk 1 starts 50 words before Chunk 0 ends — no fact falls through the cracks.

# pip install qdrant-client sentence-transformers openai
from qdrant_client import models, QdrantClient
from sentence_transformers import SentenceTransformer
from openai import OpenAI

encoder = SentenceTransformer('paraphrase-multilingual-MiniLM-L12-v2')
qdrant  = QdrantClient(":memory:")          # in-memory for demo
client  = OpenAI()
COLLECTION, DIM = "gadgets_2026", encoder.get_sentence_embedding_dimension()  # 384

# ── PHASE A: PREPARE (run once) ──────────────────────────────
def index_data(data):
    profiles = [f"{g['name']} ({g['category']}). Specs: {g['specs']}. "
                f"Pros: {g['pros']}. Cons: {g['cons']}. Price: ${g['price']}" for g in data]
    embeddings = encoder.encode(profiles)
    qdrant.create_collection(COLLECTION,
        vectors_config=models.VectorParams(size=DIM, distance=models.Distance.COSINE))
    qdrant.upload_points(COLLECTION, [
        models.PointStruct(id=i, vector=embeddings[i].tolist(), payload=data[i])
        for i in range(len(data))])

# ── PHASE B: QUERY (every request) ───────────────────────────
def retrieve(query, top_k=3):
    hits = qdrant.query_points(COLLECTION, query=encoder.encode(query).tolist(), limit=top_k)
    return [{**h.payload, "score": round(h.score, 3)} for h in hits.points]

def generate(query, context):
    ctx = "\n".join(f"- {g['name']} ({g['category']}): {g['specs']}. "
                    f"Pros: {g['pros']}. Cons: {g['cons']}. Price: ${g['price']}" for g in context)
    return client.chat.completions.create(
        model="gpt-4o-mini", temperature=0.1, max_tokens=400,
        messages=[
            {"role": "system", "content":
                "You are a 2026 AI-wearable advisor. Recommend ONLY from the products below. "
                "If none match, say so. Always mention price and one limitation. Rank multiple fits.\n\n"
                f"Available products:\n{ctx}"},
            {"role": "user", "content": query},
        ]).choices[0].message.content

def ask(query):                              # Retrieve → Augment → Generate
    return generate(query, retrieve(query, top_k=3))

You are a helpful assistant. Answer the user's question using ONLY the
provided context. If the answer is not in the context, say
"I don't have enough information to answer that." Do not use your
internal knowledge to fill in gaps.

<CONTEXT>
{{ retrieved_chunks }}
</CONTEXT>

QUESTION: {{ user_query }}

from sentence_transformers import CrossEncoder
reranker = CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')
candidates = retrieve(query, top_k=20)                       # stage 1: fast & broad
pairs = [(query, f"{g['name']}: {g['specs']} {g['pros']}") for g in candidates]
top_3 = [c for c, _ in sorted(zip(candidates, reranker.predict(pairs)),
                              key=lambda x: x[1], reverse=True)[:3]]  # stage 2: accurate

# pip install ragas
from ragas import evaluate
from ragas.metrics import faithfulness, context_relevancy, answer_relevancy

results = evaluate(dataset, metrics=[faithfulness, context_relevancy, answer_relevancy])
# → {'faithfulness': 0.95, 'context_relevancy': 0.88, 'answer_relevancy': 0.92}