File size: 8,653 Bytes
0a5e5a2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
bf6b437
 
0a5e5a2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
bf6b437
 
0a5e5a2
 
 
 
 
bf6b437
 
 
 
 
d6aba3e
bf6b437
 
 
 
 
0a5e5a2
 
 
 
bf6b437
0a5e5a2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import functools
import json
import os
import textwrap
from typing import List, Dict, Any

import gradio as gr
import requests
import torch
import torch.nn.functional as F
from transformers import AutoTokenizer, AutoModel

# -----------------------------
# Embedding utilities (from your snippet, with a couple of safety tweaks)
# -----------------------------
def last_token_pool(last_hidden_states: torch.Tensor, attention_mask: torch.Tensor) -> torch.Tensor:
    left_padding = (attention_mask[:, -1].sum() == attention_mask.shape[0])
    if left_padding:
        return last_hidden_states[:, -1]
    else:
        sequence_lengths = attention_mask.sum(dim=1) - 1
        batch_size = last_hidden_states.shape[0]
        return last_hidden_states[torch.arange(batch_size, device=last_hidden_states.device), sequence_lengths]

def get_detailed_instruct(task_description: str, query: str) -> str:
    return f"Instruct: {task_description}\nQuery: {query}"

class Qwen3Embedding:
    def __init__(self, device: str, size: str = "0.6B"):
        assert size in ["0.6B", "4B", "8B"]
        model_id = "Qwen/Qwen3-Embedding-" + size
        self.device = device

        self.tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
        # Use bfloat16 on GPU, float32 on CPU (safer on Spaces CPU)
        dtype = torch.bfloat16 if device != "cpu" else torch.float32
        self.model = AutoModel.from_pretrained(model_id, trust_remote_code=True).to(device, dtype=dtype)

        self.prompt_query = (
            "Given a natural language query, retrieve formal Rocq elements whose docstrings "
            "match the intent of the query."
        )

    @torch.inference_mode()
    def generate(self, sentence: str, is_query: bool = False) -> torch.Tensor:
        input_text = get_detailed_instruct(self.prompt_query, sentence) if is_query else sentence
        batch_dict = self.tokenizer(input_text, padding=True, truncation=True, return_tensors="pt").to(self.device)
        outputs = self.model(**batch_dict)
        embeddings = last_token_pool(outputs.last_hidden_state, batch_dict["attention_mask"])
        embeddings = F.normalize(embeddings, p=2, dim=1)
        return embeddings

    def name(self) -> str:
        return "qwen_embedding_base"

@functools.lru_cache(maxsize=3)
def get_embedder() -> Qwen3Embedding:
    return Qwen3Embedding(device="cpu", size="4B")

# -----------------------------
# Backend call
# -----------------------------
def call_retrieval_service(
    server_url: str, embedding: List[float], top_k: int, timeout: int = 60
) -> List[Dict[str, Any]]:
    if server_url.endswith("/"):
        server_url = server_url[:-1]
    url = f"{server_url}/query"

    payload = {"query": [embedding], "top_k": int(top_k)}
    resp = requests.post(url, json=payload, timeout=timeout)
    resp.raise_for_status()
    data = resp.json()
    if not isinstance(data, list):
        raise ValueError("Unexpected response format: expected a list of entries.")
    return data

# -----------------------------
# Formatting helpers
# -----------------------------
def _html_escape(s: str) -> str:
    return (
        s.replace("&", "&")
        .replace("<", "&lt;")
        .replace(">", "&gt;")
    )

def render_results(items: List[Dict[str, Any]]) -> str:
    if not items:
        return "<div>No results.</div>"

    rows = []
    for idx, it in enumerate(items, start=1):
        score = it.get("score", 0.0)
        name = it.get("name", "")
        kind = it.get("kind", "")
        doc = it.get("docstring", "") or ""
        location = it.get("location", "") or ""

        # Trim long docstrings for the summary line
        summary = " ".join(doc.strip().split())
        if len(summary) > 240:
            summary = summary[:240].rstrip() + "…"

        block = f"""
        <div class="result-card">
          <div class="header">
            <span class="rank">#{idx}</span>
            <code class="name">{_html_escape(name)}</code>
            <span class="meta">[{_html_escape(kind)}] · score={score:.4f}</span>
          </div>
          <div class="location">in {_html_escape(location)}</div>
          <details class="doc">
            <summary>{_html_escape(summary or "(no docstring)")}</summary>
            <pre>{_html_escape(doc)}</pre>
          </details>
        </div>
        """
        rows.append(block)

    style = """
    <style>
      .result-card {border: 1px solid rgba(0,0,0,.08); padding: 12px 14px; border-radius: 12px; margin-bottom: 12px;}
      .header {display:flex; gap:10px; align-items:center; flex-wrap:wrap;}
      .rank {font-weight: 700;}
      .name {font-size: 0.95rem; background: rgba(0,0,0,.03); padding: 2px 6px; border-radius: 6px;}
      .meta {opacity: 0.7;}
      .location {font-size: 0.9rem; opacity: 0.8; margin: 4px 0 8px;}
      details.doc summary {cursor: pointer; font-weight: 500; margin-bottom: 6px;}
      details.doc pre {white-space: pre-wrap; background: rgba(0,0,0,.02); padding: 10px; border-radius: 8px;}
    </style>
    """
    return style + "\n".join(rows)

# -----------------------------
# Gradio app
# -----------------------------
DEFAULT_SERVER = os.environ.get("COSIM_SERVER_URL", "https://theostos-llm4docq-cosim.hf.space")

def search(
    query: str,
    top_k: int,
    server_url: str,
    show_raw: bool,
) -> List[Any]:
    query = (query or "").strip()
    if not query:
        return [gr.update(value="<div>Please enter a query.</div>"), None]

    try:
        embedder = get_embedder()
        with torch.inference_mode():
            emb = embedder.generate(query, is_query=True)
        # Convert to plain list[float]
        emb_list = emb[0].detach().to(torch.float32).cpu().tolist()

        items = call_retrieval_service(server_url, emb_list, top_k)
        html = render_results(items)

        if show_raw:
            return [html, items]
        else:
            return [html, None]
    except requests.exceptions.RequestException as e:
        msg = f"<div style='color:#b00020'>Request error: {_html_escape(str(e))}</div>"
        return [msg, None]
    except RuntimeError as e:

        msg = f"<div style='color:#b00020'>Runtime error: {_html_escape(str(e))}</div>"
        return [msg, None]
    except Exception as e:
        msg = f"<div style='color:#b00020'>Unexpected error: {_html_escape(str(e))}</div>"
        return [msg, None]

with gr.Blocks(title="MathComp Retrieval (Qwen3 Embedding 4B)", theme=gr.themes.Soft()) as demo:
    gr.Markdown("# 🔎 MathComp Retrieval")
    status_md = gr.Markdown("⏳ Loading model… (first time may take a bit)")
    def warmup():
        try:
            _ = get_embedder()  # safe default
            return "✅ Model ready."
        except Exception as e:
            return f"⚠️ Warmup failed: {e}"

    demo.load(fn=warmup, inputs=None, outputs=status_md)
    with gr.Row():
        with gr.Column(scale=3):
            query = gr.Textbox(
                label="Query",
                placeholder="e.g., commutative group morphisms",
                lines=3,
                autofocus=True,
            )
            with gr.Row():
                top_k = gr.Slider(1, 50, value=5, step=1, label="top_k")
            with gr.Accordion("Advanced", open=False):
                server_url = gr.Textbox(value=DEFAULT_SERVER, label="Retrieval server URL")
                show_raw = gr.Checkbox(value=False, label="Also show raw JSON response")
            with gr.Row():
                run_btn = gr.Button("Search", variant="primary")
                clear_btn = gr.Button("Clear")

        with gr.Column(scale=4):
            pretty = gr.HTML(label="Results")
            raw_json = gr.JSON(label="Raw JSON", visible=False)

    def on_toggle_raw(show: bool):
        return gr.update(visible=show)

    show_raw.change(fn=on_toggle_raw, inputs=show_raw, outputs=raw_json)
    run_btn.click(
        fn=search,
        inputs=[query, top_k, server_url, show_raw],
        outputs=[pretty, raw_json],
        api_name="search",
    )
    clear_btn.click(lambda: ("", 5, "0.6B", True, DEFAULT_SERVER, False, "<div/>", None),
                    inputs=None,
                    outputs=[query, top_k, server_url, show_raw, pretty, raw_json])

    gr.Examples(
        examples=[
            ["polynomial division lemma for ringType"],
            ["matrix rank properties over finite fields"],
            ["group homomorphism kernel characterization"],
            ["bigop lemmas about summation reindexing"],
        ],
        inputs=[query],
        label="Try these",
    )

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