requirements.txt

quart>=0.20
hypercorn>=0.17
# cryptography — auto-generates TLS certificates for zero-config HTTP/2.
# Also pulls in the standard Python crypto stack (used transitively by many packages).
cryptography>=42.0
# flask + flask-compress kept for import compatibility (quart depends on flask)
flask>=3.0
flask-compress>=1.14
requests>=2.31
# urllib3 — transitive via requests; pinned explicitly to a floor that clears
# the redirect-leak CVE trio (CVE-2023-43804 / CVE-2023-45803 Cookie+body leak
# on cross-origin redirect, CVE-2024-37891 proxy-auth leak), fixed in 1.26.18/.19.
urllib3>=1.26.19
# Jinja2 — transitive via flask/quart; pinned to a floor past the sandbox/attr
# escaping bypasses (CVE-2024-22195/34064/56201/56326, CVE-2025-27516), fixed
# across 3.1.3–3.1.6. Flask 3.x already supports it.
Jinja2>=3.1.6
psutil>=5.9
# sqlite3 is built into Python — no external package needed
# tofu-search — standalone multi-engine web search + content fetch pipeline,
# extracted from this repo's former lib/search + lib/fetch. chatui wires its
# own LLM routing / browser-extension / auth-source behavior into tofu-search's
# provider seams via lib/search_bridge.py at startup.
# >=0.3.1: the browser-search fallback now relies on the BrowserProvider.fetch_html
# seam (the SERP parser lives IN the library); the bridge no longer ships its own
# parser, so an older tofu-search would leave browser-search dead.
# >=0.3.2: adds the pre-fetch relevance gate (skips fetching zero-query-overlap
# SERP junk before it costs a fetch / floods the browser bridge); default-on,
# so chatui search + Reading-Mode reports stop pulling off-topic pages.
# >=0.4.0: carries the `tofu_search.verify` subpackage (zero-LLM citation
# verifier). Below this floor, paper-mode's citation-integrity card silently
# never appears (build_citation_audit degrades to None on the missing import).
# >=0.4.1: adaptive per-engine dual-path proxy (proxied↔direct, sticky+TTL'd)
# — the bridge passes proxy_dual_attempt; below this floor that kwarg is
# rejected by an older configure(), so keep the floor at the version that
# defines it.
# >=0.4.2: docs/packaging hygiene (lxml transitive-dep note); no API change —
# floor bumped to always resolve the latest published release.
# >=0.4.3: entity-diversified rerank top-K for multi-entity comparison queries
# (search/rerank.py::_diversify_by_entity + shared authority.host_brand_labels)
# — internal reranking improvement, no public-API change.
# >=0.5.0: hard wall-clock deadlines (search_deadline_secs / fetch_url_deadline_secs).
# HARD FLOOR — not optional: lib/search_bridge.py::sync_search_config passes both
# kwargs to tofu_search.configure() UNCONDITIONALLY, and server.py calls
# install_search_bridge() at boot with no guard. An older configure() rejects the
# unknown kwargs with `TypeError: SearchConfig.__init__() got an unexpected keyword
# argument 'search_deadline_secs'`, so the server crashes on startup against any
# tofu-search < 0.5.0. Do not lower this floor without first making the bridge pass
# the deadline kwargs defensively.
# >=0.5.1: per-engine request throttle (min_request_interval_ms, default 400ms) that
# stops two concurrent search calls from self-inflicting an engine rate-limit (the
# DDG 202). SOFT floor (unlike the 0.5.0 deadline kwargs): the bridge does NOT pass
# min_request_interval_ms to configure(), so an older tofu-search does not crash — it
# just lacks the throttle. Live-tune via the TOFU_SEARCH_MIN_REQUEST_INTERVAL_MS env
# var (configure() picks it up for any field the bridge leaves unset).
tofu-search>=0.5.1
trafilatura>=1.6
# playwright is optional — powers the JS-rendered fallback in tofu_search.fetch.
# On Linux, Chromium needs X11/GTK shared libs (libatk, libgbm, libXcomposite, …).
#   With root:    sudo python -m playwright install-deps chromium
#   Without root: install.sh falls back to `conda install -c conda-forge`
#                 (atk-1.0, xorg-libxcomposite, mesa-libgbm-cos7-x86_64, …) and
#                 tofu_search/fetch/playwright_pool.py auto-augments LD_LIBRARY_PATH.
playwright>=1.40
Pillow>=10.0
python-pptx>=0.6.21
lxml>=4.9
# lxml.html.clean was extracted into a separate package in lxml 5.2+;
# justext (a trafilatura dep) still does `from lxml.html.clean import Cleaner`,
# so we must ensure the extracted package is present.
lxml_html_clean>=0.4
# BeautifulSoup — HTML fallback parser in tofu_search/fetch/html_extract.py.
# Usually pulled in transitively by trafilatura, but listed here so it's
# guaranteed even if trafilatura's deps change.
beautifulsoup4>=4.12
# python-dateutil — robust date parsing for HTML publish-date extraction
# in tofu_search/fetch/html_extract.py (imported eagerly, no ImportError fallback).
python-dateutil>=2.8
# Office document parsing — lib/doc_parser.py handles .docx/.pptx/.xlsx/
# .doc/.xls/.ppt uploads. Each has a graceful ImportError fallback, but
# shipping them by default means uploads Just Work.
python-docx>=1.0
openpyxl>=3.1
xlrd>=2.0
olefile>=0.46
mcp>=1.0
# orjson — fast C JSON encoder. REQUIRED (not optional): routes/chat.py's SSE
# state snapshot serializes the ENTIRE conversation (content+thinking+toolRounds,
# 10+ MB for long convs). Plain json.dumps holds the GIL for the whole call, so
# even via asyncio.to_thread it stalls the event loop ~40 ms and overflows the
# accept queue → the page goes dark (the 15000 incident, 2026-06). orjson encodes
# the same payload in ~5 ms, dropping the loop stall below the danger threshold.
# A stdlib iterencode fallback covers the rare inputs orjson rejects (non-str keys).
orjson>=3.9
# tiktoken — exact BPE tokenizer (cl100k/o200k) used by lib/token_counter as
# the local "exact" tier for OpenAI/Claude-family models. Without it the
# resolver chain falls through to the CJK/entropy heuristic, which under-counts
# high-entropy content (base64/minified data) and can let an oversized prompt
# slip past the proactive-compaction trigger. Optional at runtime (the
# heuristic is the always-present fallback), but shipping it makes the
# compaction token-gate substantially tighter.
tiktoken>=0.7
# PDF parsing stack — pymupdf (fitz) is used across lib/pdf_parser, routes/paper,
# and tofu_search.fetch for fast local PDF text/image extraction. pymupdf4llm is an
# optional LLM-oriented Markdown extractor built on top of pymupdf.
pymupdf>=1.24
pymupdf4llm>=0.0.17
# Markdown — server-side renderer for paper-report HTML export
# (routes/paper.py:export_report). Frontend still uses marked.js for live
# rendering; this dep is only used for downloadable .html exports.
markdown>=3.5
# PyYAML — serves the YAML variant of the OpenAPI spec
# (routes/api_docs.py::openapi_yaml). The route degrades to JSON on
# ImportError, so this is a soft feature dep; declared explicitly because it
# was previously only present transitively and went missing in clean envs.
PyYAML>=6.0
# uv powers several MCP servers (via `uvx`) — see lib/mcp/client.py. Installing
# it here guarantees uvx is on PATH inside the conda env without requiring a
# separate curl|sh bootstrap.
uv>=0.4
# SQLAlchemy Core (NOT the ORM) — used by lib/database/_core_schema.py as a
# single-source table-definition + dual-dialect DDL/DML *compiler* for NEW
# tables (PG primary, SQLite fallback), retiring the hand-maintained twin-DDL
# + regex _sql_translate path for those tables. Compile-only: no Engine is
# opened; execution stays on the existing get_db() connection. Importing the
# _core_schema module is optional — the legacy DB path does not depend on it.
sqlalchemy>=2.0
# zhconv — pure-Python (MediaWiki tables, MIT) Traditional↔Simplified Chinese
# converter. Used by lib/transcription/_zh.py as a FAIL-SAFE last gate on voice
# transcripts: Whisper-family / omni-chat ASR models emit Mandarin as Traditional
# by default, so a Simplified-Chinese user sees garbled variant text. No native
# build step (fits a vanilla self-hosted box + bootstrap.py). Soft dep at runtime:
# normalize_zh_variant() degrades to a no-op on ImportError, so the transcribe
# path works without it — shipping it makes voice input Just Work in Simplified.
# Disable/retarget via TOFU_ASR_ZH_VARIANT (default zh-cn; empty/off = raw).
zhconv>=1.4