Capstone: Mixed Folder to Gemini-Answered Queries

The whole course, end to end

This capstone combines the OCR-fallback triage from Lesson 11 with the LlamaIndex tie-in from Lesson 13, run against the entire sample_data/ folder, native-text documents and the one genuine scan together. The pipeline has exactly two stages, and only the second one touches the network:

1. Stage 1 (LiteParse, local, no network): parse every PDF fast with OCR off; for any document whose fast pass yields near-zero text, re-parse it with OCR on. Same strategy as Lesson 11, applied to build the final document set instead of just printing a table.
2. Stage 2 (LlamaIndex + Gemini, network): index all the resulting text, ask it three questions.

No document-parsing API, hosted or otherwise, appears anywhere in this pipeline. The only network calls are LlamaIndex's embedding calls and each query's LLM call, both to Gemini, both operating on plain text that already exists locally by the time they're made.

Proof the OCR fallback matters, not just a demo of it

The third question, "What time will water service be interrupted?", is only answerable because of Stage 1's OCR fallback. That fact lives exclusively in scanned_notice.pdf, a page with zero native PDF text (Lesson 8). Without the fallback, that document would contribute an empty Document to the index (Lesson 13's simpler pipeline explicitly skipped it for this reason), and this question would have nothing to retrieve a correct answer from. Watching the third answer come back correct is the capstone's actual proof that the fallback pipeline works end to end, not just that each piece works in isolation.

The code, piece by piece

def parse_folder_with_ocr_fallback(folder: Path) -> list[Document]:
    parser_no_ocr = liteparse.LiteParse(ocr_enabled=False, quiet=True)
    parser_ocr = liteparse.LiteParse(ocr_enabled=True, quiet=True)
    for pdf_path in sorted(folder.glob("*.pdf")):
        result = parser_no_ocr.parse(pdf_path)
        if len(result.text.strip()) < NEAR_ZERO_TEXT_CHARS:
            result = parser_ocr.parse(pdf_path)
        documents.append(Document(text=result.text, metadata={"source": ..., "ocr_used": ...}))

Same fallback logic as Lesson 11, now producing LlamaIndex Document objects directly, with ocr_used recorded in metadata alongside source, so it's visible afterward which documents needed the slower path.

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)

Identical to Lesson 13, just built from all 5 documents this time instead of 4.