LlamaIndex Free code examples in Python and JavaScript

llama-index-core plus whichever integrations you need. The umbrella `llama-index` package pulls in a lot you will not use; core plus two integrations is the leaner start.

pip install llama-index-core llama-index-embeddings-openai
pip install llama-index-vector-stores-qdrant

export OPENAI_API_KEY="sk-your-key"   # macOS / Linux
setx OPENAI_API_KEY "sk-your-key"     # Windows, then open a new shell

THE DOCSTORE IS THE POINT. It stores doc_id to content hash, so the second run skips every unchanged document before the embedding step — the difference between paying for the delta and paying for the archive.

from llama_index.core.ingestion import IngestionPipeline
from llama_index.core.node_parser import SentenceSplitter
from llama_index.core.storage.docstore import SimpleDocumentStore
from llama_index.embeddings.openai import OpenAIEmbedding

pipeline = IngestionPipeline(
    transformations=[
        SentenceSplitter(chunk_size=512, chunk_overlap=64),
        OpenAIEmbedding(model="text-embedding-3-small"),
    ],
    docstore=SimpleDocumentStore(),
)

nodes = pipeline.run(documents=documents)
print(f"{len(nodes)} nodes embedded")

DEDUPLICATION IS KEYED ON doc_id. A uuid4 per run means nothing ever matches and the docstore quietly does nothing — the pipeline still works, it just costs full price forever.

from llama_index.core import Document
from llama_index.core import SimpleDirectoryReader

# filename_as_id gives every file a stable id across runs.
documents = SimpleDirectoryReader(
    "./handbook", filename_as_id=True).load_data()

# Building them by hand: the id must come from the SOURCE, never
# from uuid4() or datetime.now().
documents = [
    Document(
        text=row["body"],
        doc_id=f"ticket-{row['id']}",       # stable, from the database
        metadata={"source": row["source"], "page": row["page"]},
    )
    for row in rows
]

SimpleDocumentStore lives in memory. A process that exits without persisting starts the next run with an empty map and re-embeds everything — which is the failure this whole recipe exists to avoid.

import os

PIPELINE_DIR = "./pipeline-storage"

if os.path.exists(PIPELINE_DIR):
    pipeline.load(PIPELINE_DIR)          # restore the hashes

nodes = pipeline.run(documents=documents)

pipeline.persist(PIPELINE_DIR)           # save them again

# Run it twice with no source changes: the second run should report
# zero nodes. If it does not, the ids are not stable.

The index is built from the nodes the pipeline produced. similarity_top_k is the retrieval budget — more context is not free, and past about five chunks most models get worse rather than better.

from llama_index.core import VectorStoreIndex

index = VectorStoreIndex(nodes)

engine = index.as_query_engine(similarity_top_k=5)
response = engine.query("What does error code E-4471 mean?")

print(response)

# The sources are on the response. Show them — a RAG answer without
# citations is indistinguishable from a guess.
for node in response.source_nodes:
    print(round(node.score, 4), node.metadata.get("source"))