File size: 14,856 Bytes
4c908af
 
 
 
 
 
 
 
 
db32011
4c908af
 
db32011
4c908af
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3272ef5
 
4c908af
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3272ef5
 
 
 
4c908af
 
 
 
 
 
 
 
 
 
 
 
 
 
 
db32011
4c908af
 
 
 
 
db32011
4c908af
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3272ef5
 
4c908af
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
db32011
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
from __future__ import annotations

import math
import os
import re
import time
from dataclasses import dataclass
from typing import Any, List, Optional, Sequence, Tuple

import gradio as gr
import numpy as np
import torch
from huggingface_hub import InferenceClient
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

COMPRESSION_MODEL_ID = "gravitee-io/very-small-prompt-compression"
DOWNSTREAM_MODEL = "openai/gpt-oss-20b"
EMBEDDING_MODEL = "sentence-transformers/all-MiniLM-L6-v2"
MAX_NEW_TOKENS = 96

compression_tokenizer = AutoTokenizer.from_pretrained(COMPRESSION_MODEL_ID, use_fast=True)
_MODEL_DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
compression_model = AutoModelForSeq2SeqLM.from_pretrained(COMPRESSION_MODEL_ID).to(_MODEL_DEVICE)
compression_model.eval()


@dataclass
class Segment:
    text: str
    punctuation: str


def _split_prompt(prompt: str) -> List[Segment]:
    """Split a prompt into sentence segments while retaining trailing punctuation."""
    parts = re.findall(r"[^.!?]+[.!?]*", prompt)
    segments: List[Segment] = []
    for part in parts:
        stripped = part.strip()
        if not stripped:
            continue
        punct_len = len(stripped) - len(stripped.rstrip(".?!"))
        punctuation = stripped[-punct_len:] if punct_len else ""
        content = stripped[:-punct_len].strip() if punct_len else stripped
        if content:
            segments.append(Segment(text=content, punctuation=punctuation))
    if not segments and prompt.strip():
        segments.append(Segment(text=prompt.strip(), punctuation=""))
    return segments


def _combine_segments(segments: Sequence[Segment]) -> str:
    pieces = []
    for segment in segments:
        piece = segment.text.strip()
        if segment.punctuation:
            piece = f"{piece}{segment.punctuation}"
        pieces.append(piece)
    return " ".join(piece for piece in pieces if piece).strip()


def _count_tokens(text: str) -> int:
    if not text:
        return 0
    return len(compression_tokenizer.encode(text, add_special_tokens=False))


def _call_compression_model(text: str, *, max_new_tokens: int = MAX_NEW_TOKENS) -> str:
    try:
        encoded = compression_tokenizer(text, return_tensors="pt", truncation=True, max_length=512)
        encoded = {k: v.to(_MODEL_DEVICE) for k, v in encoded.items()}
        with torch.no_grad():
            output_ids = compression_model.generate(
                **encoded,
                max_new_tokens=max_new_tokens,
                num_beams=4,
                no_repeat_ngram_size=3,
            )
        compressed = compression_tokenizer.decode(output_ids[0], skip_special_tokens=True)
    except Exception:
        return "broken"
    cleaned = compressed.strip().rstrip("?.!,;:")
    return compressed or text


def _embed(client: InferenceClient, text: str) -> Optional[np.ndarray]:
    if not text.strip():
        return None
    try:
        features = client.feature_extraction(text)
    except Exception:
        return None
    if isinstance(features, list):
        array = np.array(features[0] if features and isinstance(features[0], list) else features, dtype=np.float32)
    else:
        array = np.array(features, dtype=np.float32)
    if array.ndim == 0:
        return None
    if array.ndim > 1:
        array = array.squeeze()
    norm = np.linalg.norm(array)
    if not math.isfinite(norm) or norm == 0.0:
        return None
    return array


def _cosine_similarity(vec_a: np.ndarray | None, vec_b: np.ndarray | None) -> Optional[float]:
    if vec_a is None or vec_b is None:
        return None
    denom = float(np.linalg.norm(vec_a) * np.linalg.norm(vec_b))
    if denom == 0.0:
        return None
    return float(np.dot(vec_a, vec_b) / denom)


def _extract_text(payload: Any) -> str:
    if payload is None:
        return ""
    if isinstance(payload, str):
        return payload
    if isinstance(payload, dict):
        if "text" in payload and isinstance(payload["text"], str):
            return payload["text"]
        content = payload.get("content")
        if isinstance(content, str):
            return content
        if isinstance(content, list):
            return " ".join(_extract_text(item) for item in content)
        if content is None:
            return ""
    if isinstance(payload, list):
        return " ".join(_extract_text(item) for item in payload)
    if hasattr(payload, "content"):
        return _extract_text(getattr(payload, "content"))
    return ""


def _chat_completion(client: InferenceClient, prompt: str) -> Tuple[str, Optional[str]]:
    last_error: Optional[str] = None
    for attempt in range(2):
        try:
            completion = client.chat_completion(
                messages=[
                    {"role": "system", "content": "You are a helpful assistant. Answer concisely."},
                    {"role": "user", "content": prompt},
                ],
                max_tokens=1024,
                temperature=0.0,
                top_p=1.00,
            )
        except Exception as exc:
            last_error = f"{type(exc).__name__}: {exc}"
            continue

        try:
            choice = completion.choices[0] if completion.choices else None
            if choice is None:
                last_error = "No choices returned by downstream model."
                continue
            finish_reason = getattr(choice, "finish_reason", None)
            message = getattr(choice, "message", None)
            content = _extract_text(message)
            if not content:
                delta = getattr(choice, "delta", None)
                content = _extract_text(delta)
            if not content:
                raw_choice = getattr(choice, "content", None)
                content = _extract_text(raw_choice)
            content = content.strip()
            if content:
                return content, None
            last_error = f"Model returned an empty response (finish_reason={finish_reason})."
        except Exception as exc:
            last_error = f"{type(exc).__name__}: {exc}"
    return "", last_error or "No response generated."


def _get_client(model_id: str, token: Optional[str]) -> InferenceClient:
    return InferenceClient(model=model_id, token=token)


def _resolve_token(hf_token: Optional[str]) -> Optional[str]:
    token = (hf_token or "").strip()
    if token:
        return token
    return os.getenv("HF_TOKEN") or os.getenv("HUGGINGFACEHUB_API_TOKEN")


def compress_prompt_action(prompt: str, hf_token: Optional[str]) -> Tuple[str, str, str, str, str, str]:
    token = _resolve_token(hf_token)
    prompt = prompt.strip()
    if not prompt:
        message = "Please enter a prompt to compress."
        placeholder = "_Run **Compare Responses** after compression to see downstream outputs._"
        return ("", "", message, placeholder, placeholder, "")

    embedding_client = _get_client(EMBEDDING_MODEL, token)

    segments = _split_prompt(prompt)
    compressed_segments: List[Segment] = []
    segment_timings: List[float] = []

    for segment in segments:
        start = time.perf_counter()
        compressed_text = _call_compression_model(segment.text)
        segment_timings.append(time.perf_counter() - start)
        compressed_segments.append(Segment(text=compressed_text, punctuation=segment.punctuation))

    compressed_prompt = _combine_segments(compressed_segments).rstrip("?.!,;:")

    original_tokens = _count_tokens(prompt)
    compressed_tokens = _count_tokens(compressed_prompt)
    token_delta = original_tokens - compressed_tokens
    savings_pct = (token_delta / original_tokens * 100) if original_tokens else 0.0

    prompt_embedding_original = _embed(embedding_client, prompt)
    prompt_embedding_compressed = _embed(embedding_client, compressed_prompt)
    prompt_similarity = _cosine_similarity(prompt_embedding_original, prompt_embedding_compressed)

    prompt_metrics_lines = [
        f"**Original tokens:** {original_tokens}",
        f"**Compressed tokens:** {compressed_tokens}",
        f"**Token savings:** {token_delta} ({savings_pct:.1f}%)",
    ]
    if prompt_similarity is not None:
        prompt_metrics_lines.append(f"**Prompt cosine similarity:** {prompt_similarity:.3f}")
    if segment_timings:
        min_ms = min(segment_timings) * 1000.0
        max_ms = max(segment_timings) * 1000.0
        mean_ms = (sum(segment_timings) / len(segment_timings)) * 1000.0
        prompt_metrics_lines.append(
            f"**Segments:** {len(segment_timings)} • **Latency (ms):** min {min_ms:.1f} / mean {mean_ms:.1f} / max {max_ms:.1f}"
        )
    prompt_metrics = "<br>".join(prompt_metrics_lines)

    placeholder_response = "_Run **Compare Responses** to query the downstream model._"
    response_metrics = "Press **Compare Responses** to evaluate downstream behavior."

    return (
        prompt,
        compressed_prompt,
        prompt_metrics,
        placeholder_response,
        placeholder_response,
        response_metrics,
    )


def compare_responses_action(
    original_prompt: str, compressed_prompt: str, hf_token: Optional[str]
) -> Tuple[str, str, str, str, str, str]:
    token = _resolve_token(hf_token)
    original_prompt = (original_prompt or "").strip()
    compressed_prompt = (compressed_prompt or "").strip()

    if not original_prompt or not compressed_prompt:
        message = "Please compress a prompt before comparing responses."
        placeholder = "_No response generated._"
        return (
            original_prompt,
            compressed_prompt,
            message,
            placeholder,
            placeholder,
            "Responses unavailable.",
        )

    embedding_client = _get_client(EMBEDDING_MODEL, token)
    llm_client = _get_client(DOWNSTREAM_MODEL, token)

    original_tokens = _count_tokens(original_prompt)
    compressed_tokens = _count_tokens(compressed_prompt)
    token_delta = original_tokens - compressed_tokens
    savings_pct = (token_delta / original_tokens * 100) if original_tokens else 0.0

    prompt_embedding_original = _embed(embedding_client, original_prompt)
    prompt_embedding_compressed = _embed(embedding_client, compressed_prompt)
    prompt_similarity = _cosine_similarity(prompt_embedding_original, prompt_embedding_compressed)

    prompt_metrics_lines = [
        f"**Original tokens:** {original_tokens}",
        f"**Compressed tokens:** {compressed_tokens}",
        f"**Token savings:** {token_delta} ({savings_pct:.1f}%)",
    ]
    if prompt_similarity is not None:
        prompt_metrics_lines.append(f"**Prompt cosine similarity:** {prompt_similarity:.3f}")
    prompt_metrics = "<br>".join(prompt_metrics_lines)

    original_response, original_response_error = _chat_completion(llm_client, original_prompt)
    compressed_response, compressed_response_error = _chat_completion(llm_client, compressed_prompt)

    response_embedding_original = _embed(embedding_client, original_response)
    response_embedding_compressed = _embed(embedding_client, compressed_response)
    response_similarity = _cosine_similarity(response_embedding_original, response_embedding_compressed)

    response_metrics_lines = []
    if response_similarity is not None:
        response_metrics_lines.append(f"**Response cosine similarity:** {response_similarity:.3f}")

    original_response_display = original_response or "_No response generated for the original prompt._"
    compressed_response_display = compressed_response or "_No response generated for the compressed prompt._"
    if original_response_error:
        original_response_display += f"\n\n> {original_response_error}"
        response_metrics_lines.append("⚠️ Downstream model issue on original prompt.")
    if compressed_response_error:
        compressed_response_display += f"\n\n> {compressed_response_error}"
        response_metrics_lines.append("⚠️ Downstream model issue on compressed prompt.")

    if not response_metrics_lines:
        response_metrics_lines.append("Responses unavailable.")

    response_metrics = "<br>".join(response_metrics_lines)

    return (
        original_prompt,
        compressed_prompt,
        prompt_metrics,
        original_response_display,
        compressed_response_display,
        response_metrics,
    )


with gr.Blocks(fill_height=True, css=".gradio-container {max-width: 900px;}") as demo:
    gr.Markdown(
        """
        # Very Small Prompt Compression
        Enter a user prompt to see how the [gravitee-io/very-small-prompt-compression](https://huggingface.co/gravitee-io/very-small-prompt-compression) checkpoint trims it down,
        compares token savings, and checks semantic drift before forwarding to `openai/gpt-oss-20b`.

        Trained using the [gravitee-io/dolly-15k-prompt-compression](https://huggingface.co/datasets/gravitee-io/dolly-15k-prompt-compression) dataset.

        **Note:** Provide a Hugging Face Inference API token below (or set `HF_TOKEN`) so the demo can call the downstream model.
        """
    )

    token_input = gr.Textbox(
        label="Hugging Face token (optional)",
        type="password",
        placeholder="Paste an access token to use your own Inference quota",
    )

    prompt_input = gr.Textbox(
        label="User prompt",
        placeholder="Describe how to configure a rate limit policy in Gravitee API Management...",
        lines=4,
    )

    with gr.Row():
        compress_btn = gr.Button("Compress Prompt", variant="primary")
        compare_btn = gr.Button("Compare Responses", variant="secondary")

    original_prompt_output = gr.Textbox(label="Original prompt", lines=4, interactive=False)
    compressed_output = gr.Textbox(label="Compressed prompt", lines=4, interactive=False)
    prompt_metrics_output = gr.Markdown()

    with gr.Row():
        original_response_output = gr.Markdown(label="Response to original prompt")
        compressed_response_output = gr.Markdown(label="Response to compressed prompt")

    response_metrics_output = gr.Markdown()

    compress_btn.click(
        fn=compress_prompt_action,
        inputs=[prompt_input, token_input],
        outputs=[
            original_prompt_output,
            compressed_output,
            prompt_metrics_output,
            original_response_output,
            compressed_response_output,
            response_metrics_output,
        ],
    )

    compare_btn.click(
        fn=compare_responses_action,
        inputs=[original_prompt_output, compressed_output, token_input],
        outputs=[
            original_prompt_output,
            compressed_output,
            prompt_metrics_output,
            original_response_output,
            compressed_response_output,
            response_metrics_output,
        ],
    )


if __name__ == "__main__":
    demo.launch()