implementing-rag

name: implementing-rag description: Builds Retrieval-Augmented Generation systems using Google GenAI embeddings and Pinecone for SYD's knowledge-grounded AI features. Covers chunking, indexing, hybrid search, and evaluation. Use when implementing document Q&A, quote training from historical data, or knowledge retrieval.

RAG Implementation — SYD Stack

Tech Stack

Indexing Pipeline

from google import genai
from pinecone import Pinecone

client = genai.Client()
pc = Pinecone()
index = pc.Index("syd-knowledge")

async def index_document(doc_id: str, text: str, metadata: dict):
    # 1. Embed
    response = client.models.embed_content(
        model="models/text-embedding-004",
        contents=[text],
    )
    # 2. Upsert (idempotent via doc_id)
    index.upsert(vectors=[(doc_id, response.embeddings[0].values, metadata)])

Chunking Strategies

• Recursive splitting: Default for natural language (500 tokens, 10-15% overlap)
• Semantic chunking: Split at embedding-distance shifts for technical docs
• Markdown splitting: Preserve headers for structured renovation documents

Retrieval

# Hybrid search: semantic + keyword (SKU codes, room types)
results = index.query(
    vector=query_embedding,
    top_k=10,
    filter={"room_type": {"$eq": "bagno"}},
    include_metadata=True,
)

SYD Use Cases

1. Quote training: RAG from approved historical quotes → few-shot examples for InsightEngine
2. Price book retrieval: SKU lookup from master_price_book.json embeddings
3. Renovation knowledge: Building code references, material specifications

Best Practices

1. Idempotent indexing: Use unique document IDs to prevent duplicate chunks
2. Hybrid recall: Always combine semantic + keyword for SKU numbers and technical terms
3. Citability: Return source and page metadata in responses
4. Evaluation: Use ADK eval rubrics (see evaluating-adk-agents) to measure grounding quality