fylepy 0.1.0

Open anything, get clean Markdown for LLMs.

Any file in. Clean Markdown out. LLM ready.

A lightweight library that turns PDF, DOCX, XLSX, audio, video, and ~100 more formats into the Markdown your LLM already understands.

---

What is this?

A lightweight library for reading files. What makes it different: the output is LLM-ready — clean Markdown you can feed straight into any model, no post-processing, no cleanup.

One line. Every common file. LLM-ready Markdown. Point fyle at a path, URL, or raw bytes — what comes back is already something a model can read natively. No OCR plumbing, no format-specific parser glue, no prompt engineering to "please strip the headers and footers".

import fyle

text = fyle.read("report.pdf")   # or .docx / .xlsx / .mp3 / .mp4 / an http(s) URL / raw bytes
llm.complete(text)               # that's it.

Works out of the box on:

• PDF / DOCX / XLSX / PPTX / HTML / Markdown / CSV — parsed into Markdown
• Images — base64 data:image/... URLs ready for multimodal models
• Audio / video — local ASR transcripts with [MM:SS] timestamps (+ keyframes for video)
• SQLite — schema preview + fluent doc.table(t).query(sql) API
• Archive — safe extraction + Markdown manifest, agent decides what to open next
• ~100 source / config / log formats — passthrough as plain text

100% local. No cloud APIs. No telemetry. No API keys. Just fyle.open(...) and the file becomes something an LLM can see.

---

Install

pip install fylepy

Audio / video transcription are opt-in extras (native wheels + a ~140 MB model on first run):

pip install 'fylepy[audio]'   # faster-whisper
pip install 'fylepy[video]'   # faster-whisper + PySceneDetect + PyAV

---

Quick start

import fyle

doc = fyle.open("report.pdf")
# or: fyle.open("https://example.com/report.pdf")
# or: fyle.open(raw_bytes)   # format auto-detected from magic bytes

# Three views of the same document:
print(doc.text)            # pure content — whatever the reader produced
print(str(doc))            # LLM-ready: filename + format + size header, then content
print(repr(doc))           # short debug line for logs

# Typical usage — hand the whole thing to your model in one line:
llm.complete(str(doc))     # filename carries real signal the model can use

print(doc.meta.format)     # "pdf"
print(doc.meta.ext)        # "pdf"
print(doc.pages[0].text)   # just page 1

# One-shot convenience: str in, LLM-ready string out (same as str(fyle.open(...)))
text = fyle.read("report.pdf")

# Check which readers are available in your install
fyle.readers()
# {"pdf": ["pymupdf4llm*"], "audio": ["faster-whisper*"], ...}

---

Supported formats

Family	Extensions	Reader
PDF	.pdf	pymupdf4llm
Word	.docx	mammoth
Excel	.xlsx	openpyxl + tabulate
PowerPoint	.pptx	python-pptx
Web	.html .htm	markdownify
Markdown	.md .markdown	markdown-it-py
CSV	.csv	stdlib + tabulate
Image	.png .jpg .jpeg .webp	Pillow → base64 data URL
Audio	.mp3 .m4a .wav .flac .ogg	faster-whisper (CPU, int8)
Video	.mp4 .m4v .mov .avi .mkv .webm	PySceneDetect + Whisper
Database	.db .sqlite .sqlite3	stdlib + fluent SQL API
Archive	.zip .tar .gz .bz2 .xz ...	stdlib — extract to disk + manifest
Text	.py .js .go .rs .java .json .yaml .toml .sql .log ... (~100)	passthrough

---

Audio & video

doc = fyle.open("meeting.mp4")

print(doc.text)
# # Video: meeting.mp4
#
# - Duration: `12:34`
# - Keyframes: 8
# - Language: `en`
#
# ## Transcript
#
# [00:00] Welcome everyone to the quarterly review...

for img in doc.images:
    print(img.caption, img.src[:32])
    # "02:17"  "data:image/jpeg;base64,/9j/4AA..."

First call downloads the Whisper base model (~140 MB). CPU only — no GPU needed. Override with FYLE_WHISPER_MODEL=small (or medium / large-v3) for higher quality.

---

SQLite

doc = fyle.open("chinook.db")

for page in doc.pages:
    print(page.name)          # table or view name
    print(page.text)          # schema + sample rows

rows = doc.table("Customer").query(
    "SELECT Country, COUNT(*) AS n FROM Customer GROUP BY Country ORDER BY n DESC"
)

---

Archive

doc = fyle.open("~/Downloads/invoices.zip")

print(doc.text)                # Markdown listing of extracted files
print(doc.meta.warnings)       # ["extracted to: /.../invoices/"]

# Agent's next step: fyle.open(one of the extracted files)

Refuses .. path traversal and symlink escapes; extracts to the archive's sibling directory.

---

Chunking for RAG

for chunk in doc.chunks(max_tokens=4000, overlap=200):
    embed(chunk.text)
    # chunk.tokens / chunk.page_range also available

---

Notes

1. Offline only. Every reader runs locally. The audio/video reader downloads the Whisper model from Hugging Face on first run; after that, no network.
2. Archive reader is list-only. It extracts files to disk and returns a manifest — it does not recursively parse contents. The agent decides what to open next.
3. Alpha. Core is stable, but APIs may move between 0.x releases.

---

Feedback

Issues, PRs, and stars are welcome.

---

License

MIT © 2026 zhixiangxue

---