Why Your RAG Gives Wrong Answers — And It's Not the LLM's Fault

# New package replaces deprecated llama-parse
pip install llama-cloud-services llama-index-core

# Get your free API key at cloud.llamaindex.ai
export LLAMA_CLOUD_API_KEY="your_api_key_here"

from llama_cloud_services import LlamaParse
import os

parser = LlamaParse(
    api_key=os.environ["LLAMA_CLOUD_API_KEY"],
    result_type="markdown",       # or "json"
    verbose=True,
    language="en",
    # 🌟 Unique superpower: natural language parsing instructions
    parsing_instruction="""
        Extract all tables in full. Preserve column headers.
        Keep all numbers and units exactly as they appear.
        If a page is scanned, apply OCR carefully.
        Maintain reading order for multi-column layouts.
    """
)

# Parse async
documents = await parser.aload_data("report.pdf")
for doc in documents:
    print(doc.text)  # Clean Markdown — ready for chunking

from llama_cloud_services import LlamaParse
from llama_index.core import VectorStoreIndex, SimpleDirectoryReader
from llama_index.core.node_parser import MarkdownElementNodeParser

parser = LlamaParse(
    result_type="markdown",
    parsing_instruction="Treat all tables carefully. Preserve structure."
)

documents = SimpleDirectoryReader(
    "./docs",
    file_extractor={".pdf": parser, ".docx": parser, ".pptx": parser}
).load_data()

# Structure-aware chunking — understands Markdown headers + tables
node_parser = MarkdownElementNodeParser(num_workers=8)
nodes = node_parser.get_nodes_from_documents(documents)

index = VectorStoreIndex(nodes)
query_engine = index.as_query_engine(similarity_top_k=5)

response = query_engine.query("What was the revenue in Q3?")
print(response)

# Core install
pip install docling

# With GPU acceleration (recommended for production)
pip install docling[gpu]

# Granite-Docling-258M downloads automatically on first use
# Or pull manually from HuggingFace:
pip install huggingface-hub
python -c "from huggingface_hub import snapshot_download; \
snapshot_download('ibm-granite/granite-docling-258m')"

from docling.document_converter import DocumentConverter
from docling.datamodel.pipeline_options import PdfPipelineOptions

# Granite-Docling-258M is default VLM since January 2026
pipeline_opts = PdfPipelineOptions(
    do_ocr=True,               # OCR for scanned pages
    do_table_structure=True,   # TableFormer model
    generate_page_images=True, # Extract embedded images
)

converter = DocumentConverter(pipeline_options=pipeline_opts)
result = converter.convert("report.pdf")
doc = result.document

markdown = doc.export_to_markdown()   # Clean RAG-ready Markdown
json_data = doc.export_to_dict()       # Full DocTags structure

# Access tables as pandas DataFrames!
for table in doc.tables:
    df = table.export_to_dataframe()
    print(df)

from docling.document_converter import DocumentConverter
from docling.chunking import HybridChunker
from langchain_core.documents import Document
from langchain_openai import OpenAIEmbeddings, ChatOpenAI
from langchain_community.vectorstores import Chroma
from langchain.chains import RetrievalQA
import glob

converter = DocumentConverter()
all_chunks = []

for pdf_path in glob.glob("./docs/*.pdf"):
    result = converter.convert(pdf_path)

    # HybridChunker: structure-aware, respects headings + tables
    chunker = HybridChunker(
        tokenizer="sentence-transformers/all-MiniLM-L6-v2",
        max_tokens=512,     # 2026 sweet spot
        overlap_tokens=64,  # ~12.5% overlap
        merge_peers=True
    )

    for chunk in chunker.chunk(dl_doc=result.document):
        all_chunks.append(Document(
            page_content=chunk.text,
            metadata={
                "source":   pdf_path,
                "headings": chunk.meta.headings,
                "page":     chunk.meta.doc_items[0].prov[0].page_no
            }
        ))

vectorstore = Chroma.from_documents(all_chunks, OpenAIEmbeddings())

qa = RetrievalQA.from_chain_type(
    llm=ChatOpenAI(model="gpt-4o-mini"),
    retriever=vectorstore.as_retriever(search_kwargs={"k": 5}),
    return_source_documents=True
)
answer = qa.invoke({"query": "What are the key findings?"})
print(answer["result"])