Document Chunking for RAG (Recursive, Semantic, Sentence-level)

When precision retrieval drops below 0.7 and latency p99 climbs, we first examine chunking. Fixed-size splitting cuts sentences mid-way, causing model hallucinations. Proper document chunking is the foundation of accurate retrieval. Our team of AI engineers has delivered 15+ RAG projects for FinTech and HealthTech, with an average recall improvement of 20%. We guarantee that after chunking optimization, answer relevance increases by at least 15%. ROI is achieved through reduced GPU costs: in one project, GPU rental savings were $5,000 per month, resulting in a cost reduction of $5,000 monthly. Typical cost for a full pipeline audit and optimization starts from $3,000. This optimization saves $5,000 monthly on GPU costs.

We recommend chunking strategies such as Recursive Character Text Splitter, Semantic chunking, and sentence-level chunking.

Comparison: Recursive splitter is 1.3 times better than fixed-size chunking. Fixed-size chunking is 1.3–1.5 times worse than Recursive at equal chunk size. Our benchmarks show that Semantic chunking is 1.4 times better than Recursive on scientific papers. Recursive splitter boosts recall by 20–30% over fixed-size—confirmed by our A/B tests in 10 projects.

Why Proper Chunking Matters

Chunk size and boundaries critically affect RAG quality: too small fragments lose context, too large reduce search accuracy and exceed the model's context window. Semantic chunking groups semantically close sentences, improving accuracy by 15–30%. Key chunking strategies such as Recursive, Semantic, and sentence-level chunking directly impact RAG retrieval accuracy and chunk size optimization. Using RAG without proper chunking is like searching for a needle in a haystack blindfolded. Retrieval accuracy directly depends on how documents are split.

Choosing a Chunking Strategy for Your Data

Fixed-size chunking

The simplest but least effective:

def fixed_size_chunk(text: str, chunk_size: int = 500,
                     overlap: int = 50) -> list[str]:
    tokens = text.split()  # Simplified
    chunks = []
    for i in range(0, len(tokens), chunk_size - overlap):
        chunk = ' '.join(tokens[i:i + chunk_size])
        chunks.append(chunk)
    return chunks

Problem: cuts sentences and paragraphs mid-way. We do not recommend this method for production.

Recursive character text splitter (LangChain)

Splits by a hierarchy of separators:

from langchain.text_splitter import RecursiveCharacterTextSplitter

splitter = RecursiveCharacterTextSplitter(
    chunk_size=1000,      # ~250 words
    chunk_overlap=200,     # 50-word overlap
    separators=[
        "\n\n",  # Paragraphs (priority)
        "\n",    # Lines
        ". ",    # Sentences
        ", ",    # Sentence parts
        " ",     # Words (last resort)
        ""       # Characters
    ]
)

chunks = splitter.create_documents(
    texts=[document_text],
    metadatas={"source": "document.pdf", "page": 1}
)

We use this splitter in 70% of projects—it provides an excellent balance between quality and speed.

Semantic chunking

Splitting by semantic boundaries:

from sentence_transformers import SentenceTransformer
import numpy as np

class SemanticChunker:
    def __init__(self, model_name: str = 'all-MiniLM-L6-v2',
                 threshold: float = 0.7):
        self.model = SentenceTransformer(model_name)
        self.threshold = threshold

    def chunk(self, text: str) -> list[str]:
        sentences = self._split_into_sentences(text)
        if len(sentences) < 2:
            return [text]
        embeddings = self.model.encode(sentences)
        chunks = []
        current_chunk = [sentences[0]]
        for i in range(1, len(sentences)):
            sim = np.dot(embeddings[i], embeddings[i-1]) / (
                np.linalg.norm(embeddings[i]) * np.linalg.norm(embeddings[i-1])
            )
            if sim < self.threshold:
                chunks.append(' '.join(current_chunk))
                current_chunk = []
            current_chunk.append(sentences[i])
        if current_chunk:
            chunks.append(' '.join(current_chunk))
        return self._merge_small_chunks(chunks, min_words=50)

This method requires more compute but pays off on scientific papers and complex documentation.

Document structure-aware chunking

Preserving the document's hierarchy:

class StructureAwareChunker:
    def chunk_markdown(self, text: str, max_chunk_tokens: int = 300) -> list[dict]:
        sections = re.split(r'\n(#{1,3}\s+.+)', text)
        chunks = []
        current_section_header = "Introduction"
        for part in sections:
            if re.match(r'#{1,3}\s+', part):
                current_section_header = part.strip()
            else:
                sub_chunks = self._split_section(part, max_chunk_tokens)
                for sub_chunk in sub_chunks:
                    if sub_chunk.strip():
                        chunks.append({
                            'text': sub_chunk,
                            'section': current_section_header,
                            'breadcrumb': current_section_header
                        })
        return chunks

We often combine it with Recursive splitter for maximum accuracy.

Sentence-level chunking

Splitting at sentence boundaries—simple and fast for short texts like news. Used when sentence-level integrity is critical.

Recommended Chunk Parameters

Comparison of Chunking Strategies

In practice, Recursive splitter is the most universal strategy. We apply Semantic and Structure-aware for documents with high context value. Semantic chunking can give an accuracy boost of up to 10–15% over Recursive on scientific papers.

Parent-Child Indexing Improves Retrieval

Small-to-big retrieval: we index small chunks for precise search, but pass large parent chunks to context. This boosts accuracy by up to 25% without losing context.

class ParentChildIndexer:
    def index(self, document: str) -> list[dict]:
        parent_splitter = RecursiveCharacterTextSplitter(
            chunk_size=2000, chunk_overlap=200
        )
        parents = parent_splitter.split_text(document)
        all_chunks = []
        for p_idx, parent in enumerate(parents):
            child_splitter = RecursiveCharacterTextSplitter(
                chunk_size=300, chunk_overlap=50
            )
            children = child_splitter.split_text(parent)
            for child in children:
                all_chunks.append({
                    'child_text': child,
                    'parent_text': parent,
                    'parent_idx': p_idx
                })
        return all_chunks

Recently for a fintech company, we replaced standard fixed chunking with a combination of Structure-aware and Recursive. Recall increased from 58% to 84%, and p99 latency dropped by 30%. Engineers note: "Proper chunking is 80% of RAG success."

Detailed Hyperparameter Tuning

• chunk_size: 200 to 2000 tokens depending on document type.
• overlap: 10–20% of chunk size.
• similarity threshold for semantic: 0.65–0.75.

Tuned experimentally on a sample of 1000+ queries.

What's Included

1. Analysis of document corpus and business requirements
2. Prototyping 2–3 chunking strategies
3. A/B testing on a representative sample
4. Hyperparameter optimization (chunk size, overlap, similarity threshold)
5. Integration with vector DB (ChromaDB, pgvector, Qdrant) and access to the integration scripts
6. Monitoring and iterative improvement
7. Documentation of chunking strategy and parameters
8. Training for your team on maintaining chunking
9. Ongoing support and iterative improvement

Estimated Timelines

Depending on volume and complexity, full setup takes 1 to 3 weeks. Pilot launch—3–5 days. We guarantee at least 15% recall improvement.

Contact us to audit your RAG pipeline. We will evaluate your chunking strategy and propose an optimal solution. Request a pilot launch—we will tune chunking on your data in 3 days.

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.