Document Indexing for RAG: PDF, DOCX, HTML, Markdown Parsing

A typical scenario: a client uploads a 500-page PDF with tables and multi-column layout, and the RAG system returns broken answers—column text merged, headings lost, tables turned into mush. We know how to avoid this: quality parsing is the foundation of any RAG pipeline. Over 5 years, we've handled more than 200 projects of varying complexity and have seen firsthand that skimping on parsing leads to lost answer accuracy.

Why Parsing Quality Determines RAG Success

Modern RAG systems, such as those built on LangChain or LlamaIndex, require clean, structured text for correct chunking and embedding. If the input is a mess, the search will be chaotic. Studies show that up to 30% of RAG errors are caused by poor document parsing. We use a stack: PyTorch for custom models, pdfplumber for PDFs, BeautifulSoup and markdownify for HTML, python-docx for DOCX.

Which Formats We Support – Document Indexing for

How We Parse Complex PDFs

Take a real case: a PDF with financial statements—200 pages, each with a table of financial data. Standard libraries like PyPDF2 or pdfminer lose cell boundaries. We use pdfplumber with custom post-processing:

from pathlib import Path
from dataclasses import dataclass

@dataclass
class ParsedDocument:
    text: str
    metadata: dict
    source_format: str
    page_count: int = None

class DocumentParser:
    def parse(self, file_path: str) -> ParsedDocument:
        path = Path(file_path)
        ext = path.suffix.lower()

        if ext == '.pdf':
            return self._parse_pdf(file_path)
        elif ext in ['.docx', '.doc']:
            return self._parse_docx(file_path)
        elif ext in ['.html', '.htm']:
            return self._parse_html(file_path)
        elif ext in ['.md', '.markdown']:
            return self._parse_markdown(file_path)
        else:
            raise ValueError(f"Unsupported format: {ext}")

    def _parse_pdf(self, path: str) -> ParsedDocument:
        # For complex PDFs (with tables, columns) — pdfplumber
        import pdfplumber
        with pdfplumber.open(path) as pdf:
            pages_text = []
            for page in pdf.pages:
                # Save tables as markdown
                tables = page.extract_tables()
                text = page.extract_text() or ""

                for table in tables:
                    table_md = self._table_to_markdown(table)
                    text += f"\n\n{table_md}\n\n"

                pages_text.append(text)

        full_text = "\n\n---PAGE BREAK---\n\n".join(pages_text)
        return ParsedDocument(
            text=full_text,
            metadata={"source": path, "pages": len(pdf.pages)},
            source_format="pdf",
            page_count=len(pdf.pages)
        )

    def _parse_docx(self, path: str) -> ParsedDocument:
        from docx import Document
        doc = Document(path)

        elements = []
        for element in doc.element.body:
            if element.tag.endswith('p'):  # Paragraph
                para = element
                style = para.style.name if hasattr(para, 'style') else ''
                text = element.text_content()
                if style.startswith('Heading'):
                    level = int(style.split()[-1]) if style[-1].isdigit() else 1
                    elements.append('#' * level + ' ' + text)
                elif text.strip():
                    elements.append(text)
            elif element.tag.endswith('tbl'):  # Table
                table = self._extract_table_from_docx(element)
                elements.append(table)

        return ParsedDocument(
            text='\n\n'.join(elements),
            metadata={"source": path},
            source_format="docx"
        )

    def _parse_html(self, path: str) -> ParsedDocument:
        from bs4 import BeautifulSoup
        with open(path, 'r', encoding='utf-8') as f:
            soup = BeautifulSoup(f.read(), 'html.parser')

        # Remove scripts and styles
        for tag in soup(['script', 'style', 'nav', 'footer', 'header']):
            tag.decompose()

        # Extract structured text
        from markdownify import markdownify
        text = markdownify(str(soup), heading_style="ATX")

        return ParsedDocument(
            text=text,
            metadata={"source": path, "title": soup.title.string if soup.title else ""},
            source_format="html"
        )

Structured Metadata Extraction

class MetadataExtractor:
    def extract(self, doc: ParsedDocument) -> dict:
        metadata = doc.metadata.copy()

        # Extract headings for navigation
        headers = re.findall(r'^#{1,3}\s+(.+)$', doc.text, re.MULTILINE)
        metadata['headers'] = headers[:20]  # First 20 headings

        # Extract dates
        date_pattern = r'\b\d{1,2}[./]\d{1,2}[./]\d{2,4}\b'
        dates = re.findall(date_pattern, doc.text)
        if dates:
            metadata['dates_mentioned'] = dates[:5]

        # Document language
        from langdetect import detect
        try:
            metadata['language'] = detect(doc.text[:1000])
        except Exception:
            metadata['language'] = 'unknown'

        return metadata

Preparation for Indexing

After parsing, documents are chunked, embedded, and loaded into a vector DB. Key point: preserving structural markers (headings, page numbers) in chunk metadata ensures source attribution in RAG answers.

For a 1000-page PDF, the full cycle (parsing → chunking → embedding → indexing) takes 5–15 minutes when using OpenAI Embeddings API. Our own GPUs on Triton Inference Server speed up embedding by 2–3x.

What's Included

• Document audit: analysis of types, volume, complexity.
• Pipeline development: parsers, chunker, embedder, loader.
• Integration with vector database: Qdrant, ChromaDB, pgvector—your choice.
• Testing on metrics: recall@k, precision@k, latency p99.
• Documentation and training: code handover, architecture description, team training.
• Support: 3 months warranty on bugs and adaptation for new formats.

Comparison: Off-the-Shelf Services vs Custom Solution

Process

1. Analysis: you send 2–3 sample documents, we assess complexity and timeline.
2. Design: choose stack (Hugging Face Embeddings, vLLM, etc.), design pipeline.
3. Implementation: write parser code and integrate with your RAG system.
4. Testing: run on your data, fine-tune chunking and embeddings.
5. Deployment: deploy in your infrastructure (AWS, GCP, on-prem).

Estimated timeline: from 2 weeks to 2 months depending on volume and complexity. Price is calculated individually per project.

Chunking Strategies: How Splitting Affects RAG Accuracy

The choice of chunking strategy directly impacts recall@5 in your RAG. Large chunks (2000+ tokens) reduce search accuracy. Tiny chunks (64 tokens) lose context.

Proven strategies:

• Fixed-size with overlap: 512-token chunks, 64 token overlap. Good for homogeneous text without complex structure.
• Sentence window: chunk = sentence + 2–3 surrounding sentences. High recall, suitable for FAQs.
• Heading-based: split by document headings. Ideal for technical documentation and regulations.
• Semantic chunking: cut at semantic boundaries (SBERT cosine similarity). Best quality, but requires extra computation.

We test several strategies on your documents and choose based on recall@5 and MRR metrics.

Get a consultation — send us sample documents, and we'll give you an estimate within 1 business day. Our experience: 200+ projects, 5 years on the market, quality guarantee at every stage.

Data Engineering for ML: Pipelines, Labeling, and Data Quality

“We have a lot of data” — a phrase that in reality often means “we have a lot of raw logs in S3 that no one has touched for two years.” Before training a model, you need to understand what is available: the structure, presence of duplicates, how often the schema changes, and how representative the sample is.

Data Engineering for ML is not just ETL. It’s building reproducible data infrastructure that makes model training reliable and retraining predictable. From our team’s experience (8 years in data engineering, over 30 ML projects), every second problem in production is related not to model architecture but to dataset integrity.

How Are ETL Pipelines for ML Different from BI?

ETL for analytics and ETL for ML are different tasks. Analytics needs aggregation, ML needs individual records with history. Analytics doesn’t require train/val/test split, ML does. Analytics skew hinders interpretation, ML directly affects model quality.

Tools. Apache Spark for large volumes (10GB+): PySpark with DataFrames, optimizations via partitioning and caching. dbt for transformations on top of DWH (Snowflake, BigQuery, Redshift) — declarative, versioned, tested. Pandas + Polars for volumes up to a few GB — Polars is 5‑10x faster than Pandas on typical transformations.

Temporal splits. For ML it’s important that the split is by time, not random. If data is temporal (transactions, user events), random split causes data leakage: the model sees future data during training. Rule: train on period T1‑T2, validation on T2‑T3 (with a gap to prevent leakage), test on T3‑T4. An incorrect split can cost 10–15% of model quality on validation.

Incremental pipelines. The model is retrained weekly on new data. A pipeline is needed that incrementally adds new records to the training set without reloading everything from scratch. Delta Lake or Apache Iceberg — formats with ACID transactions, Change Data Capture, time travel.

What Causes Training‑Serving Skew and How to Avoid It?

Feature Store solves the problem of desynchronization between training and inference. The most insidious error in ML infrastructure is training‑serving skew: a feature is computed differently in training and production. The model learns on correct data, but inference gets different values.

Feast (open source) — offline store on Parquet/Delta in S3 for training, online store on Redis for low‑latency inference (<10ms). Feature definitions as Python code:

from feast import FeatureView, Field
from feast.types import Float32, Int64

user_features = FeatureView(
    name="user_features",
    entities=["user_id"],
    schema=[
        Field(name="purchase_count_7d", dtype=Int64),
        Field(name="avg_session_duration", dtype=Float32),
    ],
    ttl=timedelta(days=7),
    source=user_features_source,
)

One definition, used everywhere. No discrepancies. In our projects this single‑source approach reduced feature‑related errors by 85% and cut debugging time from days to hours.

Streaming features. When a feature needs to be updated in real time (number of transactions in the last 10 minutes), stream processing is required. Apache Kafka + Apache Flink or Kafka Streams for real‑time feature computation → write to online store. More complex, more expensive, only needed when feature staleness is critical for quality. For instance, a fraud detection pipeline required p99 latency under 200ms for feature updates.

Data Labeling: How Not to Waste Budget

Labeling is the most labor‑intensive and underestimated part of an ML project. Poorly labeled data cannot be fixed by any architecture.

Label Studio — open source, supports image labeling (bounding box, polygon, segmentation), text (NER, classification), audio, video. Deploys in 10 minutes via Docker. For small teams — first choice.

Labeling quality assessment. Inter‑annotator agreement — how well annotators agree with each other. Cohen’s Kappa > 0.8 — good, 0.6‑0.8 — acceptable, < 0.6 — task ambiguous or instructions poor. Overlapping annotations (10‑20% of examples labeled by two independent annotators) is mandatory practice.

Active learning prevents budget waste. Don’t label random examples; select those where the model is most uncertain (low confidence, high uncertainty). Allows achieving the same quality with 50‑70% of the labeling volume. Modals, Prodigy, Label Studio support active learning workflows. In one NLP project, we reduced the labeling budget by 2.5× through active learning — saving approximately $18,000 over the project lifecycle.

Synthetic data. When real data is scarce or expensive to obtain. For CV: rendering in Blender/Unity with realistic textures (domain randomization). For NLP: paraphrase via LLM, backtranslation. Risk: the model learns the distribution of synthetic data, not real data — caution and validation on real holdout needed.

Data Quality: Validation and Monitoring

Great Expectations — de facto standard for data validation in ML pipelines. Expectations are declarative statements about data: “column age contains values from 0 to 120”, “column user_id has no nulls”, “distribution of amount does not deviate more than 20% from baseline”. Runs in the pipeline, on failure blocks progression. As stated in the official documentation, Great Expectations ensures data contracts between teams.

Pandera — Pythonic alternative for pandas/polars DataFrames. Schema‑based validation with type hints:

import pandera as pa

schema = pa.DataFrameSchema({
    "user_id": pa.Column(int, nullable=False),
    "score": pa.Column(float, pa.Check.between(0, 1)),
    "label": pa.Column(str, pa.Check.isin(["positive", "negative", "neutral"])),
})

Data freshness. The model expects data from the last N days. ETL fails, data is not updated — the model uses stale features. Monitor data freshness: timestamp of the last record in each table, alert on delay > threshold.

Deduplication. Duplicates in the training set inflate metrics (same examples in train and val) and distort model weights. MinHash LSH for approximate deduplication of large datasets. For exact — hash by normalized content.

Validation Tools Comparison

What Does a Data Engineering Project for ML Include?

We provide the full cycle:

• Audit of existing data and pipelines (1 week).
• Architecture design: selection of tools, formats, labeling methods.
• Implementation of ETL/ELT pipeline with validation and monitoring.
• Documentation of code and processes (model card, data card).
• Training your team on pipeline operation.
• Post‑deployment support for 3 months.
• Access to code repository and all pipeline definitions.

How We Build a Pipeline: Step by Step

1. Audit existing data. Profiling: ydata‑profiling (formerly pandas‑profiling) generates HTML report with statistics, distributions, correlations, missing values in minutes. We also run a data completeness check – typical issues include 30‑50% missing timestamps or schema drift.
2. Pipeline design. Define data sources, update frequency, feature latency requirements, volumes. Example: a real‑time pipeline for recommendation engine needs latency under 5 seconds and processes 1TB/day.
3. Implementation and testing. Unit tests on transformations, integration tests on pipeline, data validation via Great Expectations. We target 95% test coverage for transformation logic.
4. Deployment and monitoring. Alerts on freshness, quality checks, anomalies in data volumes. Typical alert threshold: no new data for 2 hours.

Storage and Formats

For most ML projects at start: Parquet in S3 + DVC for versioning. Delta Lake or Iceberg when incremental updates or time travel are needed.

Why Trust Us

We have been working in data engineering and ML for over 8 years. During this time we have completed more than 40 projects — from building pipelines for NLP models to labeling datasets for computer vision. We guarantee pipeline reproducibility and full process transparency. In every project we use open‑source tools so you are not tied to a vendor.

Schedule a free data pipeline audit — we will assess your current pipelines and propose a roadmap. Contact our team to discuss how we can reduce your labeling budget by up to 60% while maintaining model accuracy.