Embedding Pipeline Development for Data Indexing

Without a properly built embedding pipeline, a RAG system returns irrelevant answers or loses context due to incorrect chunking. We build production-ready pipelines that ensure p99 latency under 50 ms when indexing 1 million documents. Indexing via the OpenAI API text-embedding-3-small costs around $2 per 1M documents, while self-hosted solutions on an A10G deliver nearly double the throughput and cost about $0.5 per hour of GPU. With 5+ years of experience in ML engineering and over 15 deployed semantic search systems, we guarantee reliable infrastructure.

How to Choose the Right Embedding Model?

Text chunking is a critical step. Too short chunks lose context; too long ones blur semantics. The optimal size is 256–512 tokens for English and 128–256 for Russian. We use recursive splitting with 10–20% overlap to preserve boundary meanings. Without this, even the best embedding models produce up to 30% misclicks. The choice of embedding model depends on budget, latency, and data sensitivity. The table below compares popular options:

Self-hosted models (BGE-M3) are 1.5–2 times faster than the OpenAI API on large volumes but require a dedicated GPU. For sensitive data, this is the only option.

How to Ensure Embedding Quality?

We combine several techniques: content hash deduplication to prevent double indexing, truncation of texts longer than 8000 tokens, addition of metadata (text preview and attributes) to each vector, and drift monitoring with weekly recalculation of the embedding distribution.

Pipeline Implementation

Steps to Set Up

1. Choose a model and storage: API (OpenAI, Cohere) or self-hosted (BGE-M3). Select a vector database: Qdrant, Pinecone, Weaviate.
2. Design chunking: set chunk size and overlap for your language and content type. Test on a sample of 1000 documents.
3. Implement deduplication: use SHA-256 hashing to skip duplicates during incremental indexing.
4. Batch processing: configure batch size (100 for API, 500–1000 for self-hosted) with retries using exponential backoff. According to OpenAI Embedding API best practices, a batch size of 100 is recommended.
5. Monitor and test: run load testing on 100K+ queries, measure p99 latency and error rate.

Basic Code Example

import asyncio
import hashlib
from typing import Any
import numpy as np
from openai import AsyncOpenAI
import qdrant_client
from qdrant_client.models import Distance, VectorParams, PointStruct

class EmbeddingPipeline:
    def __init__(self, collection_name: str,
                 embedding_model: str = "text-embedding-3-small"):
        self.oai = AsyncOpenAI()
        self.qdrant = qdrant_client.QdrantClient(host="localhost", port=6333)
        self.model = embedding_model
        self.collection = collection_name
        self.batch_size = 100
        self._init_collection()

    def _init_collection(self):
        dims = {"text-embedding-3-small": 1536, "text-embedding-3-large": 3072}
        dim = dims.get(self.model, 1536)
        existing = [c.name for c in self.qdrant.get_collections().collections]
        if self.collection not in existing:
            self.qdrant.create_collection(
                collection_name=self.collection,
                vectors_config=VectorParams(size=dim, distance=Distance.COSINE)
            )

    async def embed_batch(self, texts: list[str]) -> list[list[float]]:
        max_retries = 3
        for attempt in range(max_retries):
            try:
                response = await self.oai.embeddings.create(
                    input=texts,
                    model=self.model
                )
                return [item.embedding for item in response.data]
            except Exception as e:
                if attempt == max_retries - 1:
                    raise
                await asyncio.sleep(2 ** attempt)

    async def index_documents(self, documents: list[dict],
                               text_field: str = "content") -> dict:
        total = len(documents)
        indexed = 0
        skipped = 0
        existing_hashes = self._get_existing_hashes()
        new_docs = []
        for doc in documents:
            content_hash = hashlib.sha256(
                doc.get(text_field, "").encode()
            ).hexdigest()[:16]
            if content_hash not in existing_hashes:
                doc['_hash'] = content_hash
                new_docs.append(doc)
            else:
                skipped += 1
        for i in range(0, len(new_docs), self.batch_size):
            batch = new_docs[i:i + self.batch_size]
            texts = [doc.get(text_field, "") for doc in batch]
            texts = [t[:8000] for t in texts]
            embeddings = await self.embed_batch(texts)
            points = []
            for j, (doc, embedding) in enumerate(zip(batch, embeddings)):
                point_id = int(hashlib.sha256(
                    doc.get('id', str(i + j)).encode()
                ).hexdigest()[:8], 16) % (2**63)
                payload = {k: v for k, v in doc.items()
                           if k != text_field and not k.startswith('_')}
                payload['text_preview'] = doc.get(text_field, "")[:200]
                payload['_hash'] = doc['_hash']
                points.append(PointStruct(
                    id=point_id,
                    vector=embedding,
                    payload=payload
                ))
            self.qdrant.upsert(
                collection_name=self.collection,
                points=points
            )
            indexed += len(batch)
        return {'indexed': indexed, 'skipped': skipped, 'total': total}

    def _get_existing_hashes(self) -> set:
        try:
            scroll_result = self.qdrant.scroll(
                collection_name=self.collection,
                scroll_filter=None,
                with_payload=["_hash"],
                limit=10000
            )
            return {point.payload.get('_hash', '') for point in scroll_result[0]}
        except Exception:
            return set()

Advanced: Multimodal and Structured Data

class MultimodalEmbeddingPipeline:
    def __init__(self):
        self.text_pipeline = EmbeddingPipeline("text_collection")
        self.table_pipeline = EmbeddingPipeline("table_collection")

    async def process_table(self, df: pd.DataFrame,
                              table_name: str,
                              description: str = "") -> list[dict]:
        documents = []
        schema_text = f"Table: {table_name}. {description}\n"
        schema_text += f"Columns: {', '.join(df.columns.tolist())}\n"
        schema_text += f"Sample data:\n{df.head(3).to_string()}"
        documents.append({
            'id': f"{table_name}_schema",
            'content': schema_text,
            'type': 'table_schema',
            'table_name': table_name
        })
        if len(df) <= 1000:
            for idx, row in df.iterrows():
                row_text = f"Row {idx} in {table_name}: " + \
                           ", ".join([f"{col}={val}" for col, val in row.items()])
                documents.append({
                    'id': f"{table_name}_row_{idx}",
                    'content': row_text,
                    'type': 'table_row',
                    'table_name': table_name,
                    'row_index': idx
                })
        return documents

    async def process_structured_content(self, content: dict,
                                          content_type: str) -> list[dict]:
        import json
        flat_text = self._flatten_dict(content)
        return [{
            'id': content.get('id', hashlib.md5(flat_text.encode()).hexdigest()),
            'content': flat_text,
            'type': content_type,
            'raw': json.dumps(content, ensure_ascii=False)[:1000]
        }]

    def _flatten_dict(self, d: dict, prefix: str = "") -> str:
        parts = []
        for key, value in d.items():
            full_key = f"{prefix}.{key}" if prefix else key
            if isinstance(value, dict):
                parts.append(self._flatten_dict(value, full_key))
            elif isinstance(value, list):
                parts.append(f"{full_key}: {', '.join(str(v) for v in value[:10])}")
            else:
                parts.append(f"{full_key}: {value}")
        return ". ".join(parts)

Search Integration

    async def semantic_search(self, query: str,
                               limit: int = 5,
                               score_threshold: float = 0.7) -> list[dict]:
        query_embedding = (await self.text_pipeline.embed_batch([query]))[0]
        results = self.text_pipeline.qdrant.search(
            collection_name=self.text_pipeline.collection,
            query_vector=query_embedding,
            limit=limit,
            score_threshold=score_threshold,
            with_payload=True
        )
        return [
            {
                'score': hit.score,
                'text': hit.payload.get('text_preview', ''),
                'metadata': {k: v for k, v in hit.payload.items()
                             if k not in ['text_preview', '_hash']}
            }
            for hit in results
        ]

Performance Benchmarks

Incremental updating via hashing saves up to 90% of costs during daily updates. For example, total cost for 1M documents: ~$2 (API) or ~$0.5 GPU per hour for self-hosted.

Common Pitfalls

• Using an embedding model unsuitable for the data language – multilingual models (multilingual-e5-large) are needed for non-English content.
• Chunks too large (>512 tokens) reduce search accuracy by 15–20%.
• Lack of deduplication leads to duplicate vectors and storage bloat.
• Ignoring metadata – without attribute filtering, search becomes slower and less precise.

Our Expertise and Deliverables

With over 5 years of experience in ML engineering and more than 15 deployed semantic search systems, we guarantee reliable pipelines. We provide a written guarantee for p99 latency under 50 ms and cost savings up to 90% with incremental updates. Our turnkey service includes:

• Data analysis – determining optimal chunking and model.
• Architecture design – choosing a vector store, configuring batches.
• Implementation – writing production code with error handling and monitoring.
• Testing – load testing (100K+ requests) and embedding quality checks.
• Documentation and training – Model Card, operation manual, and team training.

Timeline: from 2 weeks to 2 months depending on complexity. Cost is calculated individually after an audit of your data. Contact us for a consultation and project evaluation.

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.