Reply & Attachment Extraction

Every reply carries the conversation before it: an “On Tuesday, Priya wrote:” line, the quoted history, a “Sent from my iPhone” line, an unsubscribe footer. If you pass text straight to an LLM, your agent pays for that history on every turn and can mistake an old quoted request for a new one.

AgentMail does this extraction for you. Every message has extracted_text and extracted_html alongside the raw text and html. Every attachment with extractable text has a text_url. You don’t need Talon, email-reply-parser, or any other parsing library.

Reply extraction

AgentMail computes extracted_text and extracted_html when it stores a message, for received and sent messages alike. They are returned wherever the message body is:

Given this plain-text reply:

Thursday at 2pm works. I'll send an invite.
Sent from my iPhone
On Tue, Sep 22, 2026 at 9:14 AM Priya Shah <priya@acme.com> wrote:
> Can we move the demo to Thursday?
> Anything after 1pm works for us.

extracted_text is:

Thursday at 2pm works. I'll send an invite.

Read the extracted field first, and fall back to the full body when it’s absent:

from agentmail import AgentMail
client = AgentMail(api_key="YOUR_API_KEY")
message = client.inboxes.messages.get(
    inbox_id="support@yourdomain.com",
    message_id="<CAF3x9Lq2mP7@mail.gmail.com>",
)
# new content only, falling back to the full body
content = message.extracted_text or message.text

What gets removed

• Quoted history. “On … wrote:” lines in English and many other languages, including French, Spanish, German, Portuguese, Italian, Dutch, Polish, Russian, Chinese, and Arabic. Also > quoted lines, pasted From: / Sent: / To: / Subject: header blocks, -----Original Message----- dividers, and the quote markup that Gmail, Outlook, and Apple Mail add to HTML replies.
• Trailing boilerplate. “Sent from my iPhone” and “Get Outlook for iOS” lines, unsubscribe and “manage preferences” footers, “view in browser” links, confidentiality notices, and legal disclaimers. Extraction cuts at the first such line, so everything below it is removed too.

The sender’s own signature is kept, including a -- signature that their email client placed below the quoted history.

Messages that are kept whole

Some messages contain quoted text that your agent needs, so extraction keeps them whole:

• Bounces, in both extracted_text and extracted_html. Delivery failure notices quote your original message back, and that copy shows which message failed.
• Inline replies, in extracted_text. When the sender answers between quoted lines, cutting at the first quote would lose their answers. extracted_html still cuts at the first quote, so prefer extracted_text when you pass content to a model.

When to use text and html instead

• The extracted field is absent. A message with no HTML part has no extracted_html, and event payloads for large messages leave out body fields. Fall back to the full body, as in the example above.
• You need the original. Use text or html to archive a message, to show it to a person, or when the quoted context matters to the task.
• The result looks wrong. Extraction is rule-based. An unusual layout can occasionally leave quoted text in, or remove content that follows a line that looks like boilerplate.

Building LLM context from a thread

Get Thread returns every message with its extracted content, so you can build a compact transcript without resending quoted history on every turn. Within a page, messages are ordered oldest to newest. The first page holds the newest messages, and next_page_token fetches older ones.

from agentmail import AgentMail
client = AgentMail(api_key="YOUR_API_KEY")
thread = client.inboxes.threads.get(
    inbox_id="support@yourdomain.com",
    thread_id="thread_abc123",
)
# one turn per message, new content only
transcript = "\n\n".join(
    f"{m.from_}: {m.extracted_text or m.text or ''}" for m in thread.messages
)

Token budgeting is up to you. For long threads, keep the most recent messages in full and summarize or drop older ones before sending the transcript to your model.

Attachment text extraction

AgentMail also extracts plain text from attachments, so your agent can read a PDF invoice or a spreadsheet without running a document parser. Get Attachment returns text_url, a signed URL for the extracted text, alongside download_url for the original file. Both expire at expires_at, so fetch a fresh attachment response when you need the URLs again rather than storing them.

Files larger than 10 MB are not extracted, and extracted text is cut off at 250,000 characters.

import httpx
from agentmail import AgentMail
client = AgentMail(api_key="YOUR_API_KEY")
attachment = client.inboxes.messages.get_attachment(
    inbox_id="support@yourdomain.com",
    message_id="<CAF3x9Lq2mP7@mail.gmail.com>",
    attachment_id="attach_789",
)
if attachment.text_url:
    text = httpx.get(attachment.text_url).text
else:
    # no extracted text: download the original file instead
    file_bytes = httpx.get(attachment.download_url).content

The thread-level Get Attachment endpoint returns text_url too.

Run the extractor yourself

The reply extractor that produces extracted_text and extracted_html is open source: agentextract on npm, under the MIT license. Use it to process mail that doesn’t come through AgentMail, or to check how a message will be extracted. It’s a TypeScript package built on pattern matching, with no ML model or DOM parser.

npm install agentextract

import { extractEmailBody } from "agentextract";
const { extractedText } = extractEmailBody({
  text: "Sounds good!\n\nOn Mon, Jun 1, 2026 at 9:00 AM Bob <bob@example.com> wrote:\n> old quoted message",
});
// extractedText === "Sounds good!"

The package also exports extractAttachment for attachment text. Its README includes a benchmark against Talon and TalonJS.