Do links come back?

The short answer

Not as links. The words of a link always come back. The address behind them does not come back for HTML, Word, PowerPoint or email.

PDF and Excel are different. Their link addresses come back as extra lines after the content. Nothing ties an address to its word.

A URL that is typed in the text is ordinary text. It comes back in every format.

The Markdown output has no links either. It never takes the form [text](url).

The engine is tika-server-standard 3.3.1.

Source: terraform/lambda-layers/build/versions.env (versions.env: TIKA_VERSION), accessed 2026-10-17.

What comes back, case by case

We sent each file to the production API. Every file holds the word spec as a link to https://example.com/spec. Every file also holds the URL https://example.com/plain as typed text. Every row shows its source. The inputs and the raw responses are in the repo.

What this means for your pipeline

A citation needs a page or a section. It rarely needs a link target. Read Page numbers and citations for what each format returns.

A crawler or a "see also" index needs the targets. Read them from the source file, as shown next. Keep them beside the chunk that holds the link text.

A detached PDF address is easy to mistake for content. It lands in your chunks as a bare URL line. Run the checks in Quality checks if you index it.

How to get the URLs yourself

Read the link targets from the source file. Run the snippet next to your txtfetch call. Each snippet ran on the sample file named under it.

HTML

You already hold the source. Read the href values with the standard library, next to the txtfetch call.

Python, run on links.html

from html.parser import HTMLParser


class Hrefs(HTMLParser):
    def __init__(self):
        super().__init__()
        self.found = []  # (anchor text, URL)
        self._href = None
        self._text = ""

    def handle_starttag(self, tag, attrs):
        if tag == "a":
            self._href = dict(attrs).get("href")
            self._text = ""

    def handle_data(self, data):
        if self._href:
            self._text += data

    def handle_endtag(self, tag):
        if tag == "a" and self._href:
            if self._href.startswith("http"):
                self.found.append((self._text, self._href))
            self._href = None


parser = Hrefs()
with open("links.html", encoding="utf-8") as f:
    parser.feed(f.read())
print(parser.found)

It printed

[('spec', 'https://example.com/spec')]

Word

A .docx file is a zip. The link targets sit in word/_rels/document.xml.rels. Keep the rows whose Type ends in /hyperlink.

Python, run on links.docx

import zipfile
import xml.etree.ElementTree as ET

REL = "{http://schemas.openxmlformats.org/package/2006/relationships}Relationship"

with zipfile.ZipFile("links.docx") as z:
    root = ET.fromstring(z.read("word/_rels/document.xml.rels"))

urls = [
    r.get("Target")
    for r in root.iter(REL)
    if r.get("Type", "").endswith("/hyperlink") and r.get("TargetMode") == "External"
]
print(urls)

It printed

['https://example.com/spec']

PowerPoint

python-pptx exposes the address of each text run.

Install

pip install python-pptx

Python, run on links.pptx

from pptx import Presentation

urls = []
for slide in Presentation("links.pptx").slides:
    for shape in slide.shapes:
        if not shape.has_text_frame:
            continue
        for paragraph in shape.text_frame.paragraphs:
            for run in paragraph.runs:
                if run.hyperlink.address:
                    urls.append((run.text, run.hyperlink.address))
print(urls)

It printed

[('spec', 'https://example.com/spec')]

PDF

txtfetch returns the address after the page text. To tie it to its word, read the link annotations with pypdf. Match the word by its position. This is a rough match.

Install

pip install pypdf

Python, run on links.pdf

from pypdf import PdfReader

reader = PdfReader("links.pdf")
for number, page in enumerate(reader.pages, start=1):
    for ref in page.get("/Annots", []):
        annot = ref.get_object()
        action = annot.get("/A")
        if annot.get("/Subtype") == "/Link" and action and "/URI" in action:
            x0, y0, x1, y1 = (float(n) for n in annot["/Rect"])
            words = []

            def visit(text, cm, tm, font_dict, font_size):
                x, y = tm[4], tm[5]
                if x0 - 1 <= x < x1 - 1 and y0 <= y <= y1 and text.strip():
                    words.append(text.strip())

            page.extract_text(visitor_text=visit)
            print(number, action["/URI"], words)

It printed

1 https://example.com/spec ['spec']

We have no snippet for Excel or ODT files. A scanned page has no link layer, so it has no URLs to read.

Web pages and the premium tier

We tested uploaded files. We did not test a page that txtfetch fetches from a URL. We did not measure links on the premium tier. We make no claim about them.

?quality=premium is sold as a tier, but it does not run today. VLM_ENABLED defaults to false, and no operator has switched the tier on in production. A premium request downgrades to the standard Tika path instead.

Read the HTML page, the Word page and the PDF page for the rest of each format.

Frequently asked questions

Does txtfetch keep hyperlinks?

Not as links. The link text always comes back. The target address does not come back for HTML, Word, PowerPoint or email. For PDF and Excel the address comes back as a separate line after the content.

Does format=markdown give me [text](url) links?

No. The Markdown output has headings, lists and tables. It has no links. A link in the source comes back as plain text.

Does a URL typed in the text come back?

Yes. A visible URL is ordinary text. It comes back as typed in every format.

Why does my PDF address come back after the text?

In our tests the PDF link address came back on its own line after the page text. It has no position, so nothing ties it to the word that held the link. Read the link annotations yourself if you need the pair.

How do I get link targets for citations or a crawl?

Read them from the source file next to the txtfetch call. The page shows short Python snippets for HTML, Word, PowerPoint and PDF. Each one ran on the sample files in the repo.

Does the premium tier keep links?

We did not measure links on the premium tier. We make no claim about it. The premium tier does not run today, so a premium request downgrades to the standard Tika path.