Generative AI Google Brain

gemini-use-cases-document-processing-doc-parsing-with-llamaindex-and-llamaparse

3 min read

Comparing LlamaIndex and LlamaParse for Dense Document Questioning Answering on Vertex AI

Source notebook

Repo path: gemini/use-cases/document-processing/doc_parsing_with_llamaindex_and_llamaparse.ipynb · Open on GitHub · intermediate

Compares LlamaIndex and LlamaParse RAG parsing methods for dense 10-Q document QA on Vertex AI.

Summary

This notebook teaches how to ingest, parse, index, and query a complex Alphabet 10-Q PDF using LlamaIndex and LlamaParse. It compares SimpleDirectoryReader, LangChainNodeParser, LlamaParse with SimpleDirectoryReader, and LlamaParse backed by Vertex AI Vector Search. The workflow uses Gemini on Vertex AI for answering and metadata extraction, text embeddings for retrieval, and query comparisons against known financial answers.

Key code patterns

Initialize Vertex AI and LlamaIndex models

import vertexai
vertexai.init(project=PROJECT_ID, location=LOCATION)
 
embedding_model = VertexTextEmbedding("text-embedding-005", credentials=credentials)
llm = Vertex(model="gemini-2.5-flash", temperature=0.0, max_tokens=5000)
 
Settings.embed_model = embedding_model
Settings.llm = llm

Configures LlamaIndex to use Vertex AI Gemini and Vertex text embeddings.

SimpleDirectoryReader RAG index

reader = SimpleDirectoryReader("./data")
documents = reader.load_data(show_progress=True)
 
simpledirectory_index = VectorStoreIndex.from_documents(documents)
simple_query_engine = simpledirectory_index.as_query_engine(similarity_top_k=2)

Provides the baseline LlamaIndex ingestion, indexing, and query engine path.

LangChain node parsing

parser = LangchainNodeParser(RecursiveCharacterTextSplitter())
langchain_nodes = parser.get_nodes_from_documents(documents)
 
langchainparser_index = VectorStoreIndex(nodes=langchain_nodes)
lg_query_engine = langchainparser_index.as_query_engine(similarity_top_k=2)

Shows custom chunking with LangChain before indexing nodes in LlamaIndex.

LlamaParse PDF extraction

parser = LlamaParse(
    parsing_instruction="You are a financial analyst working specifically with 10Q documents...",
    api_key="",
    result_type="text",
    language="en",
    invalidate_cache=True,
)

Uses LlamaParse with domain-specific parsing instructions for financial PDFs.

Metadata extraction and embedding

extractors = [
    QuestionsAnsweredExtractor(questions=3, llm=llm),
    KeywordExtractor(keywords=10, llm=llm),
]
pipeline = IngestionPipeline(transformations=extractors)
nodes = await pipeline.arun(documents=documents, in_place=False)
for node in nodes:
    node.embedding = embedding_model.get_text_embedding(node.get_content(metadata_mode="all"))

Adds question and keyword metadata to nodes, then embeds metadata-rich content for retrieval.

Vertex AI Vector Search store

vector_store = VertexAIVectorStore(
    project_id=PROJECT_ID,
    region=REGION,
    index_id="",
    endpoint_id="",
    gcs_bucket_name=GCS_BUCKET,
)
vector_store.add(nodes)
lp_index = VectorStoreIndex.from_vector_store(vector_store)

Persists parsed and embedded nodes into a predefined Vertex AI Vector Search index.

Models & APIs used

  • Models: gemini-2.5-flash, text-embedding-005
  • APIs / services: Vertex AI, Vertex AI Vector Search, Cloud Storage
  • SDKs / libraries: google-cloud-aiplatform, vertexai, llama-index, langchain-community, llama-index-embeddings-vertex, llama-index-llms-vertex, llama-index-core, llama_parse, google.auth, termcolor

When to use this

Use this pattern when comparing document parsing strategies for RAG over dense PDFs with complex financial tables.

Gotchas & caveats

  • Requires an initialized Google Cloud project with Vertex AI API enabled.
  • Requires a GCS bucket and a preexisting Vertex AI Vector Search index and endpoint.
  • Requires a LlamaParse API key.
  • Colab authentication is handled separately with google.colab.auth.authenticate_user().
  • The notebook refreshes Google credentials and uses a quota_project_id value in google.auth.default().
  • The LlamaParse API key, Vector Search index ID, and deployed index endpoint ID are left blank for the user to fill.
  • The notebook restarts the IPython kernel after setting up LlamaIndex model settings.

Best practices

  • Compare multiple parsing approaches against the same source document and query set.
  • Use similarity_top_k=2 consistently across query engines for apples-to-apples comparison.
  • Use domain-specific parsing instructions for LlamaParse on 10-Q financial documents.
  • Extract question and keyword metadata from parsed nodes before embedding for richer retrieval.
  • Print response source nodes, relevance scores, file names, page labels, and file paths for answer inspection.
  • Use an answer key with citation pages to evaluate generated answers.

Graph View

Backlinks