Parsing Approaches

From Text to Structure

You’ve decided to build a metadata graph — your task now is to extract sender, recipient, date, and subject from 5,000 text files.

What You’ll Learn

By the end of this lesson, you’ll be able to:

Approach 1: Parsing Libraries

Standard email formats are well-specified. Python’s email.parser handles RFC 2822 and MIME email completely — headers, body, attachments, encoding — with no pattern writing required.

import email
from email import policy

msg = email.message_from_file(
    f, policy=policy.default
)
sender = msg["From"]    # (1)
subject = msg["Subject"]
body = msg.get_body().get_content()

This works well when the format is standard. It faces limitations the moment the structure is gone — which happens as soon as you convert emails to PDFs.

Approach 2: Layout-Aware Extraction

When emails live in PDFs, the original structure (columns, tables, font sizes, bounding boxes) is encoded in the file but usually discarded during plain text extraction.

Layout-aware tools read both the text and the visual structure simultaneously — producing structured fields in a single pass rather than requiring a separate parsing step.

For email PDFs with consistent table layouts, this approach can eliminate most of the parsing challenge before it begins.

You’d do it at the same time as the extraction from PDF.

Why we didn’t show you this earlier

The most important consideration in any graph project is the graph itself — specifically, its data model.

A common first approach with many of these processes is to:

It is important to understand that Docling and other, similar packages are combining multiple jobs into a single process, and limiting some of the control you could have over them. They are easier to manage when you already understand every job they are doing.

Approach 3: Rule-Based Parsing

When the structure isn’t standard — PDF-extracted text, exported archives, proprietary formats — you encode what you know as rules. This covers a wide spectrum of sophistication:

The more structure you encode, the more your parser handles. The cost is maintenance — every new layout pattern means more rules.

This approach will only really work when your data contains predictable, repeating patterns. In such cases, you could run parsing and extraction on millions of files within a couple of hours for free.

Approach 4: ML-Based Extraction

Trained models learn to extract fields from labelled examples, rather than matching patterns you write by hand. Two approaches are relevant here:

These sit between rule-based parsing and LLMs: free at inference time, deterministic, and capable of generalising to unseen layouts.

Approach 5: LLM-Based Parsing

LLM-based parsing sends the extracted text to a language model and asks it to return structured fields. The model reads the text and infers the structure from meaning rather than patterns.

LLMs add genuine value on the hard cases — badly garbled text, unusual layouts, novel formats.

On clean, well-structured text they add cost and time but little to no improvement over cheaper alternatives.

Comparing the Approaches

Hybrid pipeline

Generally, you might opt for a hybrid pipeline that categorizes your data and reroutes each subset to the most suitable approach.

What’s Next

Over the next lessons, you’ll work through each approach in turn:

Then you’ll combine them into a hybrid pipeline, prepare the records for import, and explore the graph.