Email Filtering, Cleaning, and Vectorization for RAG

Email Filtering, Cleaning, and Vectorization for RAG

Engineers spend up to 2 hours a day searching for answers in corporate email. Email is a treasure trove of expert knowledge, but it is cluttered with quoted text, signatures, autoreplies, and spam. Without cleaning, the RAG pipeline produces garbage: recall drops to 40%. According to McKinsey, up to 28% of working time is spent on email correspondence. We have accumulated over 5 years of experience indexing email archives for large corporate clients and guarantee 95% accuracy on the test set. Savings on information search can reach 40% — a substantial amount for an average company.

The pipeline consists of three stages: filtering, cleaning, and thread reconstruction. The automated pipeline integrates into any MLOps process. The cost of indexing is calculated individually based on volume and complexity.

Connecting to Mail Servers

import imaplib
import email
from email.header import decode_header

class EmailIndexer:
    def __init__(self, imap_host: str, username: str, password: str):
        self.mail = imaplib.IMAP4_SSL(imap_host)
        self.mail.login(username, password)

    def fetch_emails(self, folder: str = "INBOX",
                     since_date: str = None,
                     max_count: int = 1000) -> list[dict]:
        self.mail.select(folder)

        search_criteria = []
        if since_date:
            search_criteria.append(f'SINCE {since_date}')

        criteria = ' '.join(search_criteria) if search_criteria else 'ALL'
        _, message_ids = self.mail.search(None, criteria)

        emails = []
        ids = message_ids[0].split()[-max_count:]

        for msg_id in ids:
            _, msg_data = self.mail.fetch(msg_id, '(RFC822)')
            msg = email.message_from_bytes(msg_data[0][1])
            parsed = self._parse_email(msg)
            if parsed:
                emails.append(parsed)

        return emails

    def _parse_email(self, msg: email.message.Message) -> dict | None:
        subject = self._decode_header(msg.get('Subject', ''))
        sender = msg.get('From', '')
        date = msg.get('Date', '')

        body = self._extract_body(msg)
        if not body or len(body.split()) < 20:
            return None

        clean_body = self._clean_email_body(body)

        return {
            'subject': subject,
            'sender': sender,
            'date': date,
            'body': clean_body,
            'thread_id': msg.get('Message-ID', ''),
            'in_reply_to': msg.get('In-Reply-To', ''),
        }

    def _clean_email_body(self, body: str) -> str:
        """Удаление quoted text, подписей, автоответов"""
        lines = body.split('\n')
        clean_lines = []

        for line in lines:
            if line.strip().startswith('>'):
                continue
            if re.match(r'^On .* wrote:$', line.strip()):
                break
            if line.strip().startswith('From:') and len(clean_lines) > 10:
                break
            clean_lines.append(line)

        text = '\n'.join(clean_lines).strip()

        signature_markers = [
            'Best regards,', 'Best,', 'Thanks,', 'Regards,',
            'С уважением,', 'Спасибо,'
        ]
        for marker in signature_markers:
            if marker in text:
                idx = text.rfind(marker)
                if len(text) - idx < 200:
                    text = text[:idx].strip()
                    break

        return text

Filtering Irrelevant Emails

After filtering and cleaning, each email is enriched with metadata: date, sender, subject, thread_id. This enables complex queries with temporal and personal filters.

class EmailRelevanceFilter:
    IGNORE_SENDERS = [
        'noreply@', 'no-reply@', 'donotreply@',
        'newsletter@', 'notifications@', 'alerts@'
    ]

    IGNORE_SUBJECT_PATTERNS = [
        r'^(Re: )?Automatic reply',
        r'^Out of (Office|office)',
        r'^Undelivered Mail Returned',
        r'^\[SPAM\]',
        r'^Meeting (invitation|canceled|accepted)',
    ]

    def is_relevant(self, email_dict: dict) -> tuple[bool, str]:
        sender = email_dict.get('sender', '').lower()
        subject = email_dict.get('subject', '')

        for ignore in self.IGNORE_SENDERS:
            if ignore in sender:
                return False, f"Auto-sender: {ignore}"

        for pattern in self.IGNORE_SUBJECT_PATTERNS:
            if re.search(pattern, subject, re.IGNORECASE):
                return False, f"System notification: {pattern}"

        if len(email_dict.get('body', '').split()) < 30:
            return False, "Body too short"

        return True, "relevant"

Thread Reconstruction

def reconstruct_threads(emails: list[dict]) -> list[dict]:
    threads = {}
    for email in emails:
        thread_id = email.get('in_reply_to') or email.get('thread_id')
        if thread_id not in threads:
            threads[thread_id] = []
        threads[thread_id].append(email)

    thread_docs = []
    for thread_id, thread_emails in threads.items():
        sorted_emails = sorted(thread_emails, key=lambda e: e.get('date', ''))
        thread_text = '\n\n---\n\n'.join([
            f"From: {e['sender']}\nDate: {e['date']}\n\n{e['body']}"
            for e in sorted_emails
        ])
        thread_docs.append({
            'thread_id': thread_id,
            'subject': sorted_emails[0]['subject'],
            'text': thread_text,
            'participants': list(set(e['sender'] for e in sorted_emails)),
            'date_range': (sorted_emails[0]['date'], sorted_emails[-1]['date'])
        })

    return thread_docs

How Thread Reconstruction Improves RAG Quality?

Indexing individual emails loses up to 40% of dialog context. Thread reconstruction assembles correspondence into coherent documents, preserving chronology and participants. This yields a recall@5 increase of 25–30% compared to flat indexing. We use the In-Reply-To field and date sorting to accurately restore chains.

Why Quoted Text Filtering Is Critical for RAG?

Quoted text occupies up to 70% of email chain volume but creates noisy embeddings during vectorization. Our regex and heuristic-based algorithm removes quotes, retaining only the original author's text. This improves answer relevance by 25–30% on the recall@5 metric. The second stage — removing signatures and autoreplies — reduces token costs by 20%.

Email Indexing Process

1. Source analysis — determine protocols (IMAP, Graph API), collect filtering requirements and legal constraints.
2. Connector setup — write integrations for Gmail/Outlook with OAuth 2.0 and certificate support.
3. Cleaning and structuring — apply filters, thread reconstruction, tag extraction.
4. Vectorization — generate embeddings (OpenAI text-embedding-3-large, 1536-dim) considering thread context. Preserve up to 85% of relevant content.
5. Storage in vector DB — load into Qdrant or pgvector with metadata (date, participants, subject).
6. RAG integration — connect semantic search over email data to your existing system.

What Is Included

• Connection to your mail server (IMAP, Graph API) with documented configuration.
• Filtering and cleaning scripts with detailed logs.
• Thread reconstruction and embedding generation.
• Vector DB deployment (Qdrant/pgvector) with index tuning.
• Semantic search integration into your RAG pipeline.
• Operations and monitoring documentation.
• Team training (2 hours remote).
• Support for incremental indexing of new emails.

Comparison of Email Cleaning Approaches

Our hybrid approach is 5 times faster than ML filtering with comparable quality: 1 second per 1000 emails with 95% accuracy on the test set.

Volume and Time Estimates

Indexing cost is calculated individually depending on volume, required cleaning level, and SLA. We guarantee indexing accuracy of at least 95% on the test set. To estimate your scenario — contact us, we'll evaluate the project in one day. Get a consultation from an engineer today.

Common Mistakes in Email Indexing

• Missing signatures and avatars — they generate extra tokens. We use a NER detector to remove them.
• Indexing without thread context — up to 40% of reply context is lost. Reconstruction is mandatory.
• Ignoring access rights — violates GDPR. We automatically apply retention policies and exclude personal emails.

Contact us — we'll help set up email indexing for your RAG tasks.

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.