ctxprune 0.2.0__tar.gz → 0.2.1__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.5
2
2
  Name: ctxprune
3
- Version: 0.2.0
3
+ Version: 0.2.1
4
4
  Summary: Context compression for AI agents: an LLMLingua-2 successor that keeps identifiers intact
5
5
  Project-URL: Homepage, https://github.com/riceharvest/ctxprune
6
6
  Project-URL: Model, https://huggingface.co/darioooooo0o/ctxprune-small
@@ -56,7 +56,9 @@ Runs on CPU in about 0.1 s per 500 tokens. Apache-2.0.
56
56
 
57
57
  At 50% of tokens kept, a reader LLM answers **51.5%** of SQuAD v2 + HotpotQA questions exactly from ctxprune's
58
58
  output vs **36.5%** from [headroom](https://github.com/headroomlabs-ai/headroom)'s Kompress-v2 (400 questions,
59
- 95% CI of the gap +10.5 to +19.5), and **68.8% vs 53.4%** for LLMLingua-2-large on our agent-text set.
59
+ 95% CI of the gap +10.5 to +19.5; the reader there is Qwen3.8-27B, which also labeled ctxprune's training data, so
60
+ this may flatter ctxprune), and **68.8% vs 53.4%** for LLMLingua-2-large on our agent-text set with an independent
61
+ reader (DeepSeek V4 Flash).
60
62
  **[Try it in your browser](https://huggingface.co/spaces/darioooooo0o/ctxprune)** (nothing is uploaded).
61
63
 
62
64
  ## Quick start
@@ -78,7 +80,7 @@ short = ctxprune.prune(tool_output, keep=0.5) # protect=True never drops IDs,
78
80
  | Shrink big Bash/MCP outputs inside **Claude Code** | `ctxprune install claude-code` | or the plugin: `/plugin install ctxprune --marketplace riceharvest/ctxprune` |
79
81
  | Compress a file or a pipe from the shell | `ctxprune FILE --keep 0.5` | `--protect`, `--stats`, reads stdin |
80
82
  | Call it from Python | `ctxprune.prune(text, keep=0.5)` | `Compressor` for batching and options |
81
- | Call it from another language / avoid the model load | `ctxprune serve` | `POST localhost:8787/compress {"text": ...}` |
83
+ | Call it from another language / avoid the model load | `ctxprune serve` | `POST localhost:8719/compress {"text": ...}` |
82
84
  | Compress retrieved documents in **LangChain** | `CtxpruneDocumentCompressor` | see below |
83
85
  | Compress retrieved nodes in **LlamaIndex** | `CtxpruneNodePostprocessor` | see below |
84
86
  | Use it inside **LLMLingua** | `PromptCompressor(model_name="darioooooo0o/ctxprune-small", use_llmlingua2=True)` | needs [LLMLingua#261](https://github.com/microsoft/LLMLingua/pull/261) |
@@ -101,7 +103,12 @@ original output. Settings: `CTXPRUNE_RATE` (0.5), `CTXPRUNE_MIN_CHARS` (6000). E
101
103
  (~2 s); run `ctxprune serve` in the background and the hook uses it instead.
102
104
 
103
105
  As a plugin, without touching settings files (needs `ctxprune` on PATH, or `uv` so it can run via `uvx`):
104
- `/plugin install ctxprune --marketplace riceharvest/ctxprune`.
106
+ `/plugin install ctxprune --marketplace riceharvest/ctxprune`. The plugin does not pre-download the model, so the
107
+ first large outputs pass through uncompressed while it downloads (~180 MB); run `ctxprune serve` once, or
108
+ `python -c "import ctxprune; ctxprune.prune('x')"`, to fetch it up front.
109
+
110
+ Saved originals in `~/.cache/ctxprune/outputs/` are raw tool output (which can include secrets) and are never
111
+ deleted automatically; clear them with `rm -rf ~/.cache/ctxprune/outputs`.
105
112
 
106
113
  ## LangChain and LlamaIndex
107
114
 
@@ -120,8 +127,8 @@ engine = index.as_query_engine(node_postprocessors=[CtxpruneNodePostprocessor(ra
120
127
  Use it where the token savings are worth that, e.g. long logs, test output, search results, retrieved docs.
121
128
  - Weak spots: source code (the teacher rarely deleted code, so the model keeps most of it) and non-English web prose,
122
129
  where LLMLingua-2-large still does better.
123
- - Our 378-question set is from our own training distribution, so the neutral SQuAD/HotpotQA comparison above is the
124
- fairer number. Per-source tables and caveats: [model card](https://huggingface.co/darioooooo0o/ctxprune-small).
130
+ - Our 378-question set is from our own training distribution, and the neutral SQuAD/HotpotQA comparison used our
131
+ labeling teacher as the reader; neither is a fully independent benchmark. Per-source tables and caveats: [model card](https://huggingface.co/darioooooo0o/ctxprune-small).
125
132
 
126
133
  By [@imdariotoo](https://x.com/imdariotoo).
127
134
 
@@ -8,7 +8,9 @@ Runs on CPU in about 0.1 s per 500 tokens. Apache-2.0.
8
8
 
9
9
  At 50% of tokens kept, a reader LLM answers **51.5%** of SQuAD v2 + HotpotQA questions exactly from ctxprune's
10
10
  output vs **36.5%** from [headroom](https://github.com/headroomlabs-ai/headroom)'s Kompress-v2 (400 questions,
11
- 95% CI of the gap +10.5 to +19.5), and **68.8% vs 53.4%** for LLMLingua-2-large on our agent-text set.
11
+ 95% CI of the gap +10.5 to +19.5; the reader there is Qwen3.8-27B, which also labeled ctxprune's training data, so
12
+ this may flatter ctxprune), and **68.8% vs 53.4%** for LLMLingua-2-large on our agent-text set with an independent
13
+ reader (DeepSeek V4 Flash).
12
14
  **[Try it in your browser](https://huggingface.co/spaces/darioooooo0o/ctxprune)** (nothing is uploaded).
13
15
 
14
16
  ## Quick start
@@ -30,7 +32,7 @@ short = ctxprune.prune(tool_output, keep=0.5) # protect=True never drops IDs,
30
32
  | Shrink big Bash/MCP outputs inside **Claude Code** | `ctxprune install claude-code` | or the plugin: `/plugin install ctxprune --marketplace riceharvest/ctxprune` |
31
33
  | Compress a file or a pipe from the shell | `ctxprune FILE --keep 0.5` | `--protect`, `--stats`, reads stdin |
32
34
  | Call it from Python | `ctxprune.prune(text, keep=0.5)` | `Compressor` for batching and options |
33
- | Call it from another language / avoid the model load | `ctxprune serve` | `POST localhost:8787/compress {"text": ...}` |
35
+ | Call it from another language / avoid the model load | `ctxprune serve` | `POST localhost:8719/compress {"text": ...}` |
34
36
  | Compress retrieved documents in **LangChain** | `CtxpruneDocumentCompressor` | see below |
35
37
  | Compress retrieved nodes in **LlamaIndex** | `CtxpruneNodePostprocessor` | see below |
36
38
  | Use it inside **LLMLingua** | `PromptCompressor(model_name="darioooooo0o/ctxprune-small", use_llmlingua2=True)` | needs [LLMLingua#261](https://github.com/microsoft/LLMLingua/pull/261) |
@@ -53,7 +55,12 @@ original output. Settings: `CTXPRUNE_RATE` (0.5), `CTXPRUNE_MIN_CHARS` (6000). E
53
55
  (~2 s); run `ctxprune serve` in the background and the hook uses it instead.
54
56
 
55
57
  As a plugin, without touching settings files (needs `ctxprune` on PATH, or `uv` so it can run via `uvx`):
56
- `/plugin install ctxprune --marketplace riceharvest/ctxprune`.
58
+ `/plugin install ctxprune --marketplace riceharvest/ctxprune`. The plugin does not pre-download the model, so the
59
+ first large outputs pass through uncompressed while it downloads (~180 MB); run `ctxprune serve` once, or
60
+ `python -c "import ctxprune; ctxprune.prune('x')"`, to fetch it up front.
61
+
62
+ Saved originals in `~/.cache/ctxprune/outputs/` are raw tool output (which can include secrets) and are never
63
+ deleted automatically; clear them with `rm -rf ~/.cache/ctxprune/outputs`.
57
64
 
58
65
  ## LangChain and LlamaIndex
59
66
 
@@ -72,8 +79,8 @@ engine = index.as_query_engine(node_postprocessors=[CtxpruneNodePostprocessor(ra
72
79
  Use it where the token savings are worth that, e.g. long logs, test output, search results, retrieved docs.
73
80
  - Weak spots: source code (the teacher rarely deleted code, so the model keeps most of it) and non-English web prose,
74
81
  where LLMLingua-2-large still does better.
75
- - Our 378-question set is from our own training distribution, so the neutral SQuAD/HotpotQA comparison above is the
76
- fairer number. Per-source tables and caveats: [model card](https://huggingface.co/darioooooo0o/ctxprune-small).
82
+ - Our 378-question set is from our own training distribution, and the neutral SQuAD/HotpotQA comparison used our
83
+ labeling teacher as the reader; neither is a fully independent benchmark. Per-source tables and caveats: [model card](https://huggingface.co/darioooooo0o/ctxprune-small).
77
84
 
78
85
  By [@imdariotoo](https://x.com/imdariotoo).
79
86
 
@@ -8,7 +8,7 @@
8
8
  from .compress import DEFAULT_MODEL, Compressor
9
9
 
10
10
  __all__ = ["Compressor", "prune", "DEFAULT_MODEL"]
11
- __version__ = "0.2.0"
11
+ __version__ = "0.2.1"
12
12
 
13
13
  _default: Compressor | None = None
14
14
 
@@ -3,7 +3,7 @@
3
3
  ctxprune build.log # keep ~50% of the text, print it
4
4
  cat build.log | ctxprune --keep 0.33 --protect --stats
5
5
  ctxprune install claude-code # compress big Bash/MCP outputs inside Claude Code
6
- ctxprune serve # local HTTP API on :8787 (skips the model load per call)
6
+ ctxprune serve # local HTTP API on :8719 (skips the model load per call)
7
7
  """
8
8
 
9
9
  from __future__ import annotations
@@ -17,7 +17,7 @@ USAGE = """\
17
17
  ctxprune FILE [--keep 0.5] [--protect] [--stats] compress a file (or stdin) and print it
18
18
  ctxprune install claude-code [--scope user|project|local]
19
19
  ctxprune uninstall claude-code [--scope ...]
20
- ctxprune serve [--port 8787] [--host 127.0.0.1]
20
+ ctxprune serve [--port 8719] [--host 127.0.0.1]
21
21
  ctxprune hook claude-code (called by Claude Code, reads JSON on stdin)"""
22
22
 
23
23
 
@@ -96,7 +96,7 @@ def main(argv: list[str] | None = None) -> None:
96
96
  if cmd == "serve":
97
97
  ap = argparse.ArgumentParser(prog="ctxprune serve")
98
98
  ap.add_argument("--host", default="127.0.0.1")
99
- ap.add_argument("--port", type=int, default=8787)
99
+ ap.add_argument("--port", type=int, default=8719)
100
100
  ap.add_argument("--model", default=DEFAULT_MODEL)
101
101
  ap.add_argument("--fp32", action="store_true")
102
102
  args = ap.parse_args(argv[1:])
@@ -15,7 +15,7 @@ What it does, and what it deliberately leaves alone:
15
15
  - Any error means no output: Claude then gets the original text (fail open).
16
16
 
17
17
  Settings (environment of the Claude Code process): CTXPRUNE_RATE (default 0.5),
18
- CTXPRUNE_MIN_CHARS (6000), CTXPRUNE_SERVER (default http://127.0.0.1:8787; used when
18
+ CTXPRUNE_MIN_CHARS (6000), CTXPRUNE_SERVER (default http://127.0.0.1:8719; used when
19
19
  `ctxprune serve` is running, which avoids the ~2 s model load per compressed call).
20
20
  """
21
21
 
@@ -51,7 +51,7 @@ def _save_original(text: str) -> str:
51
51
 
52
52
  def _compress(texts: list[str], rate: float) -> list[str]:
53
53
  """Compress via a running `ctxprune serve` if there is one, else in process."""
54
- server = os.environ.get("CTXPRUNE_SERVER", "http://127.0.0.1:8787")
54
+ server = os.environ.get("CTXPRUNE_SERVER", "http://127.0.0.1:8719")
55
55
  if server:
56
56
  try:
57
57
  import urllib.request
@@ -61,9 +61,11 @@ def _compress(texts: list[str], rate: float) -> list[str]:
61
61
  data=json.dumps({"texts": texts, "rate": rate,
62
62
  "force_protected": True}).encode())
63
63
  with urllib.request.urlopen(req, timeout=60) as r:
64
- return [o["text"] for o in json.load(r)["results"]]
65
- except OSError:
66
- pass # not running: load the model here
64
+ outs = [o["text"] for o in json.load(r)["results"]]
65
+ if len(outs) == len(texts) and all(isinstance(o, str) for o in outs):
66
+ return outs
67
+ except Exception:
68
+ pass # not running, or something else owns the port: load the model here
67
69
  from .compress import Compressor
68
70
 
69
71
  c = Compressor()
@@ -1,7 +1,7 @@
1
1
  """Small local HTTP API so other processes (the Claude Code hook, Node/Go/shell tools) skip the model load.
2
2
 
3
- ctxprune serve # http://127.0.0.1:8787
4
- curl -s localhost:8787/compress -d '{"text": "...", "rate": 0.5}'
3
+ ctxprune serve # http://127.0.0.1:8719
4
+ curl -s localhost:8719/compress -d '{"text": "...", "rate": 0.5}'
5
5
 
6
6
  POST /compress {"text": str} or {"texts": [str]}, optional "rate" (0.5), "force_protected" (false),
7
7
  "keep_structure" (false) -> {"text", "kept", "total"} or {"results": [...]}
@@ -43,7 +43,7 @@ class _State:
43
43
  return out
44
44
 
45
45
 
46
- def make_server(host: str = "127.0.0.1", port: int = 8787, model: str = DEFAULT_MODEL,
46
+ def make_server(host: str = "127.0.0.1", port: int = 8719, model: str = DEFAULT_MODEL,
47
47
  onnx_file: str = INT8) -> ThreadingHTTPServer:
48
48
  state = _State(model, onnx_file)
49
49
 
@@ -1,6 +1,6 @@
1
1
  [project]
2
2
  name = "ctxprune"
3
- version = "0.2.0"
3
+ version = "0.2.1"
4
4
  description = "Context compression for AI agents: an LLMLingua-2 successor that keeps identifiers intact"
5
5
  readme = "README.md"
6
6
  license = "Apache-2.0"
@@ -7,7 +7,7 @@ sdk: static
7
7
  app_file: index.html
8
8
  pinned: false
9
9
  license: apache-2.0
10
- short_description: Agent context compression vs LLMLingua-2, real outputs
10
+ short_description: Paste text, see what ctxprune keeps (runs in your browser)
11
11
  models:
12
12
  - darioooooo0o/ctxprune-small
13
13
  - microsoft/llmlingua-2-bert-base-multilingual-cased-meetingbank
@@ -15,3 +15,6 @@ models:
15
15
 
16
16
  Real side-by-side outputs of [ctxprune-small](https://huggingface.co/darioooooo0o/ctxprune-small) and LLMLingua-2.
17
17
  Generated by `space/static/build.py` in [riceharvest/ctxprune](https://github.com/riceharvest/ctxprune).
18
+
19
+ The "Try it" box runs the int8 model in your browser with transformers.js (`js/ctxprune.js` is a port of the
20
+ Python package's atomizer and selection). The side-by-side section shows precomputed outputs.
@@ -65,3 +65,29 @@ def test_hook_fails_open(monkeypatch, capsys):
65
65
  monkeypatch.setattr("sys.stdin", io.StringIO("not json"))
66
66
  hook.main_claude_code()
67
67
  assert capsys.readouterr().out == ""
68
+
69
+
70
+ def test_foreign_server_on_port_falls_back(monkeypatch):
71
+ """Something other than `ctxprune serve` answering on the port must not break compression."""
72
+ import io
73
+ import urllib.request
74
+
75
+ class Resp(io.BytesIO):
76
+ def __enter__(self):
77
+ return self
78
+
79
+ def __exit__(self, *a):
80
+ return False
81
+
82
+ monkeypatch.setattr(urllib.request, "urlopen", lambda *a, **k: Resp(b'{"object": "not ctxprune"}'))
83
+ called = {}
84
+
85
+ class FakeCompressor:
86
+ def compress(self, t, rate, force_protected):
87
+ called["yes"] = True
88
+ return {"text": t[:10]}
89
+
90
+ import ctxprune.compress
91
+
92
+ monkeypatch.setattr(ctxprune.compress, "Compressor", lambda *a, **k: FakeCompressor())
93
+ assert hook._compress(["x" * 100], 0.5) == ["x" * 10] and called
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes