Guide: Get LlamaIndex and Pinecone working together fast.

I see too many teams get bogged down trying to wire up LlamaIndex and Pinecone for a simple RAG proof of concept. They get lost in abstractions before they've even defined their SLI for retrieval latency. Let's fix that.

Here's the fastest path to a working state. We'll define an SLO for retrieval accuracy later, but first, get the pipeline running.

**Core Dependencies**
You need these. Pin the versions to avoid surprise breaking changes.
```bash
pip install llama-index==0.10.0 llama-index-vector-stores-pinecone==0.10.0 pinecone-client==3.2.2
```

**The Minimal Setup Script**
This assumes you have your Pinecone API key and environment ready. No fancy service contexts, no custom embeddings to start.

```python
import os
from llama_index.core import VectorStoreIndex, SimpleDirectoryReader
from llama_index.vector_stores.pinecone import PineconeVectorStore
from pinecone import Pinecone

# Set your SLIs later. For now, just get the keys.
os.environ['PINECONE_API_KEY'] = 'your-api-key'

# Initialize Pinecone client
pc = Pinecone(api_key=os.environ['PINECONE_API_KEY'])
index_name = "llamaindex-quickstart"
# Use a small dimension for the default service's embedding model
dimension = 1536
metric = "cosine"
spec = {"serverless": {"cloud": "aws", "region": "us-east-1"}}

# Create the index if it doesn't exist
if index_name not in pc.list_indexes().names():
pc.create_index(name=index_name, dimension=dimension, metric=metric, spec=spec)

# Connect the vector store
pinecone_index = pc.Index(index_name)
vector_store = PineconeVectorStore(pinecone_index=pinecone_index)

# Ingest some documents
documents = SimpleDirectoryReader("./data").load_data()
index = VectorStoreIndex.from_documents(documents, vector_store=vector_store)

# Query it
query_engine = index.as_query_engine()
response = query_engine.query("What is the main idea?")
print(response)
```

**Common Pitfalls & SRE Notes**
* **Index Dimension Mismatch**: Your vector store dimension must match your embedding model's output. The default in LlamaIndex uses `text-embedding-ada-002` (1536). If you change the embedding service, your SLO on accuracy will break.
* **Namespace Confusion**: Pinecone uses namespaces for multi-tenancy. The default is an empty string `""`. Track your usage per namespace for cost and performance SLIs.
* **Metadata Bloat**: Pinecone has a 40KB metadata limit per vector. LlamaIndex can serialize metadata into the text store if you exceed this, but it's a reliability risk. Define a schema early.
* **The SLA Lie**: Pinecone's SLA is for uptime, not query p99 latency. You need to define and measure your own retrieval latency SLI. Start with a simple `vector_store.query` duration metric.

Get this running first. Then, instrument the retrieval step, log your queries, and start thinking about your actual SLOs for recall and latency. Without those, you're just building a demo.