dethrottled 0.1.1__tar.gz

dethrottled-0.1.1/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 dethrottled contributors
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

dethrottled-0.1.1/PKG-INFO

@@ -0,0 +1,267 @@
+Metadata-Version: 2.4
+Name: dethrottled
+Version: 0.1.1
+Summary: Zero-API web search, fetch and extraction. No keys, no quota, no metered pool.
+License: MIT
+Project-URL: Homepage, https://github.com/zataraine/dethrottled
+Project-URL: Repository, https://github.com/zataraine/dethrottled
+Project-URL: Issues, https://github.com/zataraine/dethrottled/issues
+Keywords: search,scraping,extraction,rag,web,no-api-key
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Topic :: Internet :: WWW/HTTP :: Indexing/Search
+Classifier: Topic :: Text Processing :: Markup :: HTML
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: requests>=2.31
+Requires-Dist: feedparser>=6.0
+Requires-Dist: ddgs>=9.0
+Requires-Dist: trafilatura>=1.8
+Requires-Dist: resiliparse>=1.0
+Requires-Dist: selectolax>=0.3.21
+Requires-Dist: fastapi>=0.110
+Requires-Dist: uvicorn>=0.27
+Provides-Extra: documents
+Requires-Dist: pymupdf>=1.24; extra == "documents"
+Requires-Dist: openpyxl>=3.1; extra == "documents"
+Requires-Dist: python-docx>=1.1; extra == "documents"
+Requires-Dist: python-pptx>=0.6.23; extra == "documents"
+Requires-Dist: xlrd>=2.0; extra == "documents"
+Requires-Dist: odfpy>=1.4; extra == "documents"
+Requires-Dist: ebooklib>=0.18; extra == "documents"
+Requires-Dist: striprtf>=0.0.26; extra == "documents"
+Provides-Extra: media
+Requires-Dist: youtube-transcript-api>=1.0; extra == "media"
+Provides-Extra: semantic
+Requires-Dist: numpy>=1.24; extra == "semantic"
+Requires-Dist: onnxruntime>=1.17; extra == "semantic"
+Requires-Dist: transformers>=4.40; extra == "semantic"
+Provides-Extra: rerank
+Requires-Dist: flashrank>=0.2.9; extra == "rerank"
+Provides-Extra: tls
+Requires-Dist: curl_cffi>=0.7; extra == "tls"
+Provides-Extra: all
+Requires-Dist: dethrottled[documents,media,rerank,semantic,tls]; extra == "all"
+Provides-Extra: dev
+Requires-Dist: pytest>=8.0; extra == "dev"
+Requires-Dist: httpx>=0.27; extra == "dev"
+Requires-Dist: ruff>=0.4; extra == "dev"
+Dynamic: license-file
+
+# dethrottled
+
+**Web search, fetch and extraction with no API keys, no accounts, and no quota.**
+
+Ask it a question and it returns results. Give it a URL and it returns the text
+— from a web page, a PDF, a spreadsheet, a slide deck, an ebook, or a video's
+captions. It runs on your machine, and there is nothing to sign up for.
+
+```bash
+pip install 'dethrottled[all]'
+dethrottled
+```
+
+```bash
+curl -X POST localhost:8787/search-and-fetch \
+  -H 'content-type: application/json' \
+  -d '{"query": "okapi bm25 ranking", "num_results": 3}'
+```
+
+Or bring up the whole self-hosted stack — search engines, a JavaScript renderer
+and the API — with nothing leaving your network:
+
+```bash
+docker compose up -d
+```
+
+> Want the whole story? [**TLDREADME.md**](TLDREADME.md) documents every tier,
+> every measurement, and every decision in full. This page is the short version.
+
+---
+
+## Why it exists
+
+Most search and scraping tools ask for an API key, meter you, and cut you off.
+The ones that don't usually fall over on the pages that matter — the JavaScript
+app, the PDF with no text layer, the spreadsheet where the actual numbers live.
+
+dethrottled handles those, locally, for nothing. Every number below is measured
+and reproducible from scripts in this repository.
+
+## Three endpoints
+
+| endpoint | what it does |
+| --- | --- |
+| `POST /search` | a query → ranked results |
+| `POST /fetch` | URLs → their text (`raw: true` for the source) |
+| `POST /search-and-fetch` | both in one call, fetching only the winners |
+
+Plus `/corpus/search` (query what you've already fetched, no network),
+`/health`, `/v2/status`, `/v2/capabilities` and `/stats`.
+
+There is deliberately **no `/crawl`**. Rendering is a strategy for obtaining a
+page, not something a caller wants for its own sake — the ladder escalates by
+itself, and a caller choosing the renderer by hand would spend four seconds of
+Chromium on a page `direct` serves in two hundred milliseconds. Where forcing
+it is genuinely useful, that's the `render` parameter.
+
+`/extract` and `/search-and-extract` remain as aliases.
+
+## The fetch ladder
+
+Three tiers, tried in order, and **a tier only succeeds if it yields readable
+prose**. That's the whole design, and it comes from a measurement:
+
+```
+page          direct   tls    crawl4ai
+quotes-js          0     0       ✓        ← JavaScript-only page
+indeed             0  2340       ✓        ← 403s an ordinary TLS handshake
+wikipedia      26861 26861       ✓
+```
+
+`quotes-js` returns **HTTP 200, a complete response body, and zero characters
+of article text.** A ladder that escalates on failed fetches declares victory
+there and hands you an empty shell. So escalation is driven by recovered text,
+and anything under 600 characters counts as a miss.
+
+| tier | speed | what it's for | needs |
+| --- | --- | --- | --- |
+| `direct` | ~2.1s | most pages. requests + trafilatura | nothing |
+| `tls` | ~0.3s | a real Chrome TLS fingerprint, no browser | a library |
+| `crawl4ai` | ~4.4s | renders JavaScript, locally | a container you run |
+
+Only `direct` is required. `tls` is **faster than plain requests** (310ms vs
+536ms median) because curl-impersonate is C — it's not a slow fallback, it's a
+quick one. A shorter ladder solves fewer pages and nothing breaks.
+
+There is deliberately **no relay tier**. Every option needs an account (a
+Cloudflare Worker), a third party who then learns every URL you fetch (a hosted
+reader, a CORS proxy), or an address range that anti-bot vendors blocklist on
+sight (Tor publishes its exit list in real time).
+
+## Everything it can read
+
+Content type is decided from the **file signature**, never the server's header —
+an HTML error page served as `application/vnd.ms-excel` is routine, and handing
+that to a spreadsheet parser produces either a crash or, worse, nonsense that
+looks like data.
+
+| kind | formats |
+| --- | --- |
+| web | HTML, via trafilatura → resiliparse → selectolax |
+| documents | PDF (+ OCR for scans), DOCX, XLSX, XLS, PPTX, CSV/TSV |
+| open formats | ODT, ODS, ODP, EPUB, RTF |
+| video | YouTube captions, from the URL |
+
+Routing costs **0.6µs** for the common case. MarkItDown does this same job with
+a neural network; this is forty lines of byte comparison.
+
+## Ranking
+
+Results arrive from three sources in no order at all. Up to three stages fix
+that, each optional:
+
+1. **BM25** — lexical, no model, microseconds
+2. **Corpus merge** — passages you've already fetched, competing with the web
+3. **Cross-encoder** — a real model, over a shortlist of 40 only
+
+Measured on a pool where one document is right and the rest are plausible
+neighbours: **MRR 0.646 → 0.833**. It costs 16ms and earns it.
+
+Two orderings are deliberate: **rank before fetching**, because fetching is the
+expensive step; and **merge the corpus before ranking**, so the reranker judges
+web and local results on equal terms.
+
+## The corpus
+
+Every page fetched is split into passages, embedded and stored in SQLite, so
+the cache quietly becomes searchable:
+
+```bash
+curl -G localhost:8787/corpus/search --data-urlencode 'q=how are documents ranked'
+```
+
+No vector database. Brute-force cosine over 50,000 passages takes **2.5ms** —
+an index would be machinery with nothing to do.
+
+## It learns which sites are readable
+
+Some domains reliably return nothing: single-page apps that render from a
+private API, syndication aggregators, consent walls. Ranking answers *"is this
+relevant?"* and has no idea whether a page can be **read**.
+
+So dethrottled records per-domain extraction outcomes and uses them at exactly
+one point — choosing which of the already-ranked results to spend a fetch on.
+It never reorders by relevance and never drops a result.
+
+Crucially it's **measured, not declared**. No blocklist ships. Evidence decays
+with a 14-day half-life, so a site that starts working recovers on its own.
+
+## Install
+
+```bash
+pip install 'dethrottled[all]'           # everything below. Start here
+```
+
+The base install is deliberately small — eight pure-Python dependencies — and
+gives you search, the `direct` fetch tier and the extraction cascade. Every
+other capability is an extra, because the heavy pieces are the ones many
+callers never use:
+
+```bash
+pip install dethrottled                  # search, direct fetch, extraction
+pip install 'dethrottled[documents]'     # + PDF, Office, ODF, EPUB, RTF, OCR
+pip install 'dethrottled[tls]'           # + the TLS tier
+pip install 'dethrottled[semantic]'      # + the corpus
+pip install 'dethrottled[rerank]'        # + the cross-encoder
+pip install 'dethrottled[media]'         # + video transcripts
+```
+
+Every optional import is guarded: a missing extra removes a capability, it does
+not break one, and `/v2/capabilities` reports exactly what this install can
+actually do rather than what the code is capable of.
+
+Runs on **x86-64 and ARM64**, Linux, macOS and Windows, Python 3.10–3.13. CI
+tests both architectures. Nothing needs a compiler — **including on a Pi**.
+
+One model, 87MB:
+
+```bash
+./scripts/fetch-models.sh
+```
+
+## Politeness
+
+This fetches other people's pages, so: `robots.txt` honoured at every tier and
+cached; one request per domain at a time with a 1.5s floor; an honest,
+contactable User-Agent; bounded retries; a 10MB response ceiling.
+
+That 1.5s floor means **fetching many URLs from one site is slow by design** —
+8 URLs from one domain take 12.7s, 8 URLs from eight domains take 3.3s. That's
+the politeness working, not a defect.
+
+## Licensing
+
+MIT, and it ships no weights. Everything it downloads is permissive:
+all-MiniLM-L6-v2 (Apache-2.0) and ms-marco-MiniLM-L-12-v2 (Apache-2.0).
+
+There is deliberately **no non-commercially-licensed component in any code
+path** — "optional" is not something a licence audit can rely on.
+
+## Documentation
+
+- [**TLDREADME.md**](TLDREADME.md) — everything, in depth. Every tier, every
+  measurement, every rejected alternative
+- [USAGE.md](USAGE.md) — the API and every configuration knob
+- [ARCHITECTURE.md](ARCHITECTURE.md) — how the pieces fit
+- [SECURITY.md](SECURITY.md) — **read before exposing this to a network**
+
+## Licence
+
+MIT. See [LICENSE](LICENSE).

dethrottled-0.1.1/README.md

@@ -0,0 +1,211 @@
+# dethrottled
+
+**Web search, fetch and extraction with no API keys, no accounts, and no quota.**
+
+Ask it a question and it returns results. Give it a URL and it returns the text
+— from a web page, a PDF, a spreadsheet, a slide deck, an ebook, or a video's
+captions. It runs on your machine, and there is nothing to sign up for.
+
+```bash
+pip install 'dethrottled[all]'
+dethrottled
+```
+
+```bash
+curl -X POST localhost:8787/search-and-fetch \
+  -H 'content-type: application/json' \
+  -d '{"query": "okapi bm25 ranking", "num_results": 3}'
+```
+
+Or bring up the whole self-hosted stack — search engines, a JavaScript renderer
+and the API — with nothing leaving your network:
+
+```bash
+docker compose up -d
+```
+
+> Want the whole story? [**TLDREADME.md**](TLDREADME.md) documents every tier,
+> every measurement, and every decision in full. This page is the short version.
+
+---
+
+## Why it exists
+
+Most search and scraping tools ask for an API key, meter you, and cut you off.
+The ones that don't usually fall over on the pages that matter — the JavaScript
+app, the PDF with no text layer, the spreadsheet where the actual numbers live.
+
+dethrottled handles those, locally, for nothing. Every number below is measured
+and reproducible from scripts in this repository.
+
+## Three endpoints
+
+| endpoint | what it does |
+| --- | --- |
+| `POST /search` | a query → ranked results |
+| `POST /fetch` | URLs → their text (`raw: true` for the source) |
+| `POST /search-and-fetch` | both in one call, fetching only the winners |
+
+Plus `/corpus/search` (query what you've already fetched, no network),
+`/health`, `/v2/status`, `/v2/capabilities` and `/stats`.
+
+There is deliberately **no `/crawl`**. Rendering is a strategy for obtaining a
+page, not something a caller wants for its own sake — the ladder escalates by
+itself, and a caller choosing the renderer by hand would spend four seconds of
+Chromium on a page `direct` serves in two hundred milliseconds. Where forcing
+it is genuinely useful, that's the `render` parameter.
+
+`/extract` and `/search-and-extract` remain as aliases.
+
+## The fetch ladder
+
+Three tiers, tried in order, and **a tier only succeeds if it yields readable
+prose**. That's the whole design, and it comes from a measurement:
+
+```
+page          direct   tls    crawl4ai
+quotes-js          0     0       ✓        ← JavaScript-only page
+indeed             0  2340       ✓        ← 403s an ordinary TLS handshake
+wikipedia      26861 26861       ✓
+```
+
+`quotes-js` returns **HTTP 200, a complete response body, and zero characters
+of article text.** A ladder that escalates on failed fetches declares victory
+there and hands you an empty shell. So escalation is driven by recovered text,
+and anything under 600 characters counts as a miss.
+
+| tier | speed | what it's for | needs |
+| --- | --- | --- | --- |
+| `direct` | ~2.1s | most pages. requests + trafilatura | nothing |
+| `tls` | ~0.3s | a real Chrome TLS fingerprint, no browser | a library |
+| `crawl4ai` | ~4.4s | renders JavaScript, locally | a container you run |
+
+Only `direct` is required. `tls` is **faster than plain requests** (310ms vs
+536ms median) because curl-impersonate is C — it's not a slow fallback, it's a
+quick one. A shorter ladder solves fewer pages and nothing breaks.
+
+There is deliberately **no relay tier**. Every option needs an account (a
+Cloudflare Worker), a third party who then learns every URL you fetch (a hosted
+reader, a CORS proxy), or an address range that anti-bot vendors blocklist on
+sight (Tor publishes its exit list in real time).
+
+## Everything it can read
+
+Content type is decided from the **file signature**, never the server's header —
+an HTML error page served as `application/vnd.ms-excel` is routine, and handing
+that to a spreadsheet parser produces either a crash or, worse, nonsense that
+looks like data.
+
+| kind | formats |
+| --- | --- |
+| web | HTML, via trafilatura → resiliparse → selectolax |
+| documents | PDF (+ OCR for scans), DOCX, XLSX, XLS, PPTX, CSV/TSV |
+| open formats | ODT, ODS, ODP, EPUB, RTF |
+| video | YouTube captions, from the URL |
+
+Routing costs **0.6µs** for the common case. MarkItDown does this same job with
+a neural network; this is forty lines of byte comparison.
+
+## Ranking
+
+Results arrive from three sources in no order at all. Up to three stages fix
+that, each optional:
+
+1. **BM25** — lexical, no model, microseconds
+2. **Corpus merge** — passages you've already fetched, competing with the web
+3. **Cross-encoder** — a real model, over a shortlist of 40 only
+
+Measured on a pool where one document is right and the rest are plausible
+neighbours: **MRR 0.646 → 0.833**. It costs 16ms and earns it.
+
+Two orderings are deliberate: **rank before fetching**, because fetching is the
+expensive step; and **merge the corpus before ranking**, so the reranker judges
+web and local results on equal terms.
+
+## The corpus
+
+Every page fetched is split into passages, embedded and stored in SQLite, so
+the cache quietly becomes searchable:
+
+```bash
+curl -G localhost:8787/corpus/search --data-urlencode 'q=how are documents ranked'
+```
+
+No vector database. Brute-force cosine over 50,000 passages takes **2.5ms** —
+an index would be machinery with nothing to do.
+
+## It learns which sites are readable
+
+Some domains reliably return nothing: single-page apps that render from a
+private API, syndication aggregators, consent walls. Ranking answers *"is this
+relevant?"* and has no idea whether a page can be **read**.
+
+So dethrottled records per-domain extraction outcomes and uses them at exactly
+one point — choosing which of the already-ranked results to spend a fetch on.
+It never reorders by relevance and never drops a result.
+
+Crucially it's **measured, not declared**. No blocklist ships. Evidence decays
+with a 14-day half-life, so a site that starts working recovers on its own.
+
+## Install
+
+```bash
+pip install 'dethrottled[all]'           # everything below. Start here
+```
+
+The base install is deliberately small — eight pure-Python dependencies — and
+gives you search, the `direct` fetch tier and the extraction cascade. Every
+other capability is an extra, because the heavy pieces are the ones many
+callers never use:
+
+```bash
+pip install dethrottled                  # search, direct fetch, extraction
+pip install 'dethrottled[documents]'     # + PDF, Office, ODF, EPUB, RTF, OCR
+pip install 'dethrottled[tls]'           # + the TLS tier
+pip install 'dethrottled[semantic]'      # + the corpus
+pip install 'dethrottled[rerank]'        # + the cross-encoder
+pip install 'dethrottled[media]'         # + video transcripts
+```
+
+Every optional import is guarded: a missing extra removes a capability, it does
+not break one, and `/v2/capabilities` reports exactly what this install can
+actually do rather than what the code is capable of.
+
+Runs on **x86-64 and ARM64**, Linux, macOS and Windows, Python 3.10–3.13. CI
+tests both architectures. Nothing needs a compiler — **including on a Pi**.
+
+One model, 87MB:
+
+```bash
+./scripts/fetch-models.sh
+```
+
+## Politeness
+
+This fetches other people's pages, so: `robots.txt` honoured at every tier and
+cached; one request per domain at a time with a 1.5s floor; an honest,
+contactable User-Agent; bounded retries; a 10MB response ceiling.
+
+That 1.5s floor means **fetching many URLs from one site is slow by design** —
+8 URLs from one domain take 12.7s, 8 URLs from eight domains take 3.3s. That's
+the politeness working, not a defect.
+
+## Licensing
+
+MIT, and it ships no weights. Everything it downloads is permissive:
+all-MiniLM-L6-v2 (Apache-2.0) and ms-marco-MiniLM-L-12-v2 (Apache-2.0).
+
+There is deliberately **no non-commercially-licensed component in any code
+path** — "optional" is not something a licence audit can rely on.
+
+## Documentation
+
+- [**TLDREADME.md**](TLDREADME.md) — everything, in depth. Every tier, every
+  measurement, every rejected alternative
+- [USAGE.md](USAGE.md) — the API and every configuration knob
+- [ARCHITECTURE.md](ARCHITECTURE.md) — how the pieces fit
+- [SECURITY.md](SECURITY.md) — **read before exposing this to a network**
+
+## Licence
+
+MIT. See [LICENSE](LICENSE).

dethrottled-0.1.1/pyproject.toml

@@ -0,0 +1,138 @@
+[build-system]
+requires = ["setuptools>=68"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "dethrottled"
+version = "0.1.1"
+description = "Zero-API web search, fetch and extraction. No keys, no quota, no metered pool."
+readme = "README.md"
+license = { text = "MIT" }
+requires-python = ">=3.10"
+keywords = ["search", "scraping", "extraction", "rag", "web", "no-api-key"]
+classifiers = [
+    "Development Status :: 4 - Beta",
+    "Intended Audience :: Developers",
+    "License :: OSI Approved :: MIT License",
+    "Programming Language :: Python :: 3.10",
+    "Programming Language :: Python :: 3.11",
+    "Programming Language :: Python :: 3.12",
+    "Programming Language :: Python :: 3.13",
+    "Topic :: Internet :: WWW/HTTP :: Indexing/Search",
+    "Topic :: Text Processing :: Markup :: HTML",
+]
+
+# The base install fetches, extracts and searches. Everything past that is an
+# extra, because the heavy pieces are exactly the ones most callers never use:
+# onnxruntime and transformers are a few hundred megabytes to make a feature
+# optional, and pymupdf is not free either. Every optional import in the source
+# is guarded, so an absent extra removes a capability rather than breaking one.
+dependencies = [
+    "requests>=2.31",
+    "feedparser>=6.0",
+    # Search: one pure-Python dependency covering a dozen keyless engines.
+    "ddgs>=9.0",
+    # Extraction cascade, in order. All three ship aarch64 wheels, so a
+    # Raspberry Pi installs them rather than compiling them.
+    "trafilatura>=1.8",
+    "resiliparse>=1.0",
+    "selectolax>=0.3.21",
+    "fastapi>=0.110",
+    "uvicorn>=0.27",
+]
+
+[project.urls]
+Homepage = "https://github.com/zataraine/dethrottled"
+Repository = "https://github.com/zataraine/dethrottled"
+Issues = "https://github.com/zataraine/dethrottled/issues"
+
+[project.optional-dependencies]
+# Documents as text: PDF, the Office formats, OpenDocument, EPUB and RTF.
+#
+# Every one of these is a direct parser. MarkItDown was measured as the
+# alternative and rejected: `markitdown[all]` is 375MB across 52 packages and
+# pulls in four Azure SDKs and a neural network (magika) to identify file
+# types -- in a tool whose entire claim is that it needs no account and no key.
+# It also wraps these same libraries, so it would be an abstraction layer over
+# dependencies we already have. File typing here is forty lines of signature
+# comparison and costs under a microsecond on the common path.
+documents = [
+    "pymupdf>=1.24",
+    "openpyxl>=3.1",
+    "python-docx>=1.1",
+    "python-pptx>=0.6.23",
+    "xlrd>=2.0",
+    # OpenDocument. Pure Python (an sdist, but nothing to compile) and what
+    # most European public bodies publish in.
+    "odfpy>=1.4",
+    # An EPUB is a zip of XHTML, read with the same extractor as a web page.
+    "ebooklib>=0.18",
+    # RTF, tiny and pure Python.
+    "striprtf>=0.0.26",
+]
+
+# Video pages, read as their caption track.
+#
+# Keyless and accountless, but it talks to YouTube's internal endpoint and
+# YouTube blocks datacentre address ranges hard -- so this works from a machine
+# you own and often will not from a cloud host. Separated from `documents` so
+# that limitation is an explicit choice rather than something that arrives
+# with a spreadsheet parser.
+media = ["youtube-transcript-api>=1.0"]
+# The semantic corpus and the multilingual reranker. Also needs weights --
+# see scripts/fetch-models.sh.
+semantic = [
+    "numpy>=1.24",
+    "onnxruntime>=1.17",
+    "transformers>=4.40",
+]
+# The English cross-encoder. flashrank downloads its own model on first use.
+rerank = ["flashrank>=0.2.9"]
+# A real Chrome TLS fingerprint without a browser. Worth its place: it is the
+# only thing that got past one host's blocking in testing, and it is FASTER
+# than plain requests, not slower. aarch64 wheels available.
+tls = ["curl_cffi>=0.7"]
+all = ["dethrottled[documents,media,semantic,rerank,tls]"]
+dev = ["pytest>=8.0", "httpx>=0.27", "ruff>=0.4"]
+
+[project.scripts]
+dethrottled = "dethrottled.server:main"
+dethrottled-health = "dethrottled.health:main"
+
+[tool.setuptools.packages.find]
+where = ["src"]
+
+[tool.pytest.ini_options]
+testpaths = ["tests"]
+markers = [
+    "network: hits the live internet; excluded from CI with -m 'not network'",
+]
+
+[tool.ruff]
+# 100, not 88. The bulk of this code sits comfortably under 80; the exceptions
+# are all format strings that read worse broken across two lines than they do
+# slightly wide. Picking a limit the code already meets beats reformatting
+# working code to satisfy a number.
+line-length = 100
+target-version = "py310"
+
+[tool.ruff.lint]
+select = ["E", "W", "F", "I", "B", "PIE", "RUF"]
+ignore = [
+    # This codebase uses %-formatting throughout, deliberately and
+    # consistently. Rewriting 52 call sites to f-strings would be churn, and a
+    # half-converted file is worse than either style applied whole.
+    "UP031",
+    # Broad excepts are the design here, not an oversight: every tier, every
+    # extractor and every optional model is meant to fail soft and degrade to
+    # the next thing. They are individually commented where the reason is not
+    # obvious.
+    "BLE001", "S110", "S112",
+    # zip() without strict= is fine where the two sequences are built together
+    # three lines above.
+    "B905",
+]
+
+[tool.ruff.lint.per-file-ignores]
+# Tests unpack tuples they only half-use, and say so.
+"tests/*" = ["RUF059"]

dethrottled-0.1.1/setup.cfg

@@ -0,0 +1,4 @@
+[egg_info]
+tag_build = 
+tag_date = 0
+

dethrottled-0.1.1/src/dethrottled/__init__.py

@@ -0,0 +1,2 @@
+"""Zero-API web search, fetch and extraction."""
+__version__ = "0.1.1"

dethrottled-0.1.1/src/dethrottled/__main__.py

@@ -0,0 +1,4 @@
+from .server import main
+
+if __name__ == "__main__":
+    main()