Feeding LiteParse into LlamaIndex

Soft prerequisite

This lesson assumes basic familiarity with the LlamaIndex course, specifically Document (Lesson 2) and VectorStoreIndex (Lesson 4). If those are new, a quick read through that course's first four lessons will make this one click faster; this lesson doesn't re-explain what a vector index is.

Where the network call actually happens

This pipeline has two distinct halves that are easy to blur together:

1. Parsing (LiteParse): entirely local, no network call, covered in every earlier lesson in this course.
2. Indexing and querying (LlamaIndex + Gemini): this is where the network enters, once, for embedding the text and once per query for the LLM's answer.

The PDF never leaves your machine. Only the extracted text (already plain strings by the time LlamaIndex sees them) goes to Gemini, for embeddings and for answering questions, the exact opposite of a hosted parsing API where the raw document itself gets uploaded.

The code, piece by piece

parser = liteparse.LiteParse(ocr_enabled=False, quiet=True)
for pdf_path in pdf_paths:
    result = parser.parse(pdf_path)
    documents.append(Document(text=result.text, metadata={"source": pdf_path.name}))

Each PDF becomes one Document. metadata={"source": ...} carries the originating filename through the index, so later a query result can be traced back to which file it came from, the whole-document equivalent of the per-page numbers Lesson 4 introduced.

scanned_notice.pdf is deliberately excluded from this lesson's batch: with ocr_enabled=False its .text comes back empty (Lesson 8), and an empty Document contributes nothing but noise to the index. Lesson 14's capstone handles the OCR-fallback case properly, this lesson keeps the pipeline simple and focuses on the LlamaIndex tie-in itself.

Settings.llm = GoogleGenAI(model="gemini-3.5-flash-lite", api_key=API_KEY)
Settings.embed_model = GoogleGenAIEmbedding(model_name="models/gemini-embedding-001", api_key=API_KEY)
index = VectorStoreIndex.from_documents(documents)
query_engine = index.as_query_engine()

Standard LlamaIndex setup (LlamaIndex course, Lessons 1-4): configure the global Settings, build a VectorStoreIndex from the Document list, get a query engine.

response = query_engine.query(question)
sources = {node.metadata.get("source") for node in response.source_nodes}

Each query retrieves relevant chunks across ALL indexed documents (not just one), and response.source_nodes reports which document(s) the answer was actually retrieved from, using the metadata["source"] LiteParse's output was tagged with earlier. This lesson's two questions each target a different source document, confirming the index is doing real cross-document retrieval, not just echoing whichever document happened to be indexed first.