File size: 9,224 Bytes
7113658
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Task 4 — Agent-Agent Single-Round Communication.

Agent A (Designer) sees a patent figure and describes it using a structured schema.
Agent B (Maker) receives the description and selects the correct figure from a pool.

Conditions tested:
  same_model  — both agents use the same model (e.g., both Gemma 4)
  cross_model — Agent A and B use different models (tests if description is grounded
                in the drawing or in model-specific vocabulary)

The structured schema forces Agent A to produce parseable output:
  {viewpoint, visible_components, spatial_relations, distinguishing_features}

This directly tests the McCarthy/mrCAD finding that design intent communication
requires both drawing (viewpoint, spatial_relations) and text (component vocabulary).

Usage:
    python scripts/eval/agent_agent.py \
        --enriched data/enriched/enriched_2022.parquet \
        --n 18 \
        --out results/agent_agent_results.json
"""

import argparse
import json
import re
import time
from collections import defaultdict
from pathlib import Path

import pandas as pd

import sys
sys.path.insert(0, str(Path(__file__).parent))
from provider import chat, get_client, get_model


# ── Schema ────────────────────────────────────────────────────────────────────

DESIGNER_PROMPT = """You are an engineer describing a technical patent figure to a colleague who cannot see it.

Patent title: {title}
This figure shows FIG. {fig_num} of the patent.

Describe the figure using ONLY this JSON schema — be specific enough that someone could identify this exact figure from among similar drawings:

{{
  "viewpoint": "...",
  "visible_components": ["...", "..."],
  "spatial_relations": "...",
  "distinguishing_features": "..."
}}

Rules:
- viewpoint: name the projection type (e.g. "front elevational view", "perspective view")
- visible_components: list 2-5 distinct parts visible in this specific view
- spatial_relations: describe how parts relate spatially (left/right/center/above/below)
- distinguishing_features: what makes THIS view unique vs. other views of the same object

Reply with the JSON only, no other text."""


MAKER_PROMPT = """You are selecting a patent figure based on a colleague's description.

The description is:
{description}

Below are {n} candidate figure descriptions (from the SAME patent but different views).
Which candidate BEST matches the description above?

{candidates}

Reply with just the number (1-{n})."""


# ── Viewpoint parsing (shared) ────────────────────────────────────────────────

def parse_vp(desc: str, fig_num: int) -> str:
    pat = re.compile(
        rf"FIG\.\s*{fig_num+1}\s+is\s+(?:a\s+|an\s+)?(.{{5,80}}?)\s*(?:view|thereof|;|\n|$)",
        re.IGNORECASE,
    )
    m = pat.search(desc or "")
    return m.group(1).strip().lower() if m else ""


def find_eligible_patents(df: pd.DataFrame, n: int, seed: int = 42) -> list[str]:
    df["vp"] = df.apply(lambda r: parse_vp(r.get("drawing_description", ""), r["figure_number"]), axis=1)
    has_persp = df.groupby("patent_id")["vp"].apply(lambda x: x.str.contains("perspective").any())
    has_front = df.groupby("patent_id")["vp"].apply(
        lambda x: (x.str.contains("front elev") | x.isin(["front elevation", "front plan"])).any()
    )
    eligible = has_persp.index[has_persp & has_front].tolist()
    import random
    rng = random.Random(seed)
    rng.shuffle(eligible)
    return eligible[:n]


# ── Agent calls ───────────────────────────────────────────────────────────────

def agent_a_describe(client, title: str, fig_num: int) -> dict | None:
    """Agent A: produce a structured description of the figure (text only — no image)."""
    prompt = DESIGNER_PROMPT.format(title=title, fig_num=fig_num + 1)
    msgs = [{"role": "user", "content": prompt}]
    response = chat(client, msgs, max_tokens=300)
    try:
        m = re.search(r"\{.*\}", response, re.DOTALL)
        if m:
            return json.loads(m.group())
    except Exception:
        pass
    return None


def agent_b_select(client, description: dict, candidates: list[dict]) -> int:
    """Agent B: select which candidate matches the description. Returns 0-indexed."""
    desc_str = json.dumps(description, indent=2)
    cand_lines = []
    for i, c in enumerate(candidates, 1):
        cand_lines.append(f"{i}. {json.dumps(c)}")
    cands_str = "\n".join(cand_lines)
    prompt = MAKER_PROMPT.format(
        description=desc_str,
        n=len(candidates),
        candidates=cands_str,
    )
    msgs = [{"role": "user", "content": prompt}]
    response = chat(client, msgs, max_tokens=10)
    m = re.search(r"\b([1-9])\b", response)
    return int(m.group()) - 1 if m else -1


# ── Main eval ─────────────────────────────────────────────────────────────────

def run(enriched_path: str, n: int, out_path: str, pool_size: int = 4, seed: int = 42):
    import random
    client = get_client()

    df = pd.read_parquet(enriched_path)
    eval_pids = find_eligible_patents(df, n, seed)
    print(f"Evaluating {len(eval_pids)} patents | model: {get_model()}")

    results = []
    correct = total = 0
    rng = random.Random(seed)

    for pid in eval_pids:
        group = df[df["patent_id"] == pid].sort_values("figure_number")
        figs = group.to_dict("records")
        title = str(group["patent_title"].iloc[0])

        # Pick a target figure (front elevation if available, else random non-perspective)
        front = [r for r in figs if "front elev" in r.get("vp", "")]
        target = front[0] if front else rng.choice([r for r in figs if "perspective" not in r.get("vp", "")] or figs)
        fig_num = int(target["figure_number"])

        # Agent A: describe the target figure (text only — uses drawing_description context)
        description = agent_a_describe(client, title, fig_num)
        if not description:
            continue
        time.sleep(0.5)

        # Build candidates: target + (pool_size-1) other views from the same patent
        others = [r for r in figs if r["figure_number"] != fig_num]
        rng.shuffle(others)
        distractor_figs = others[:pool_size - 1]

        # Describe distractors too (Agent A perspective)
        candidates_data = []
        correct_pos = rng.randint(0, pool_size - 1)
        for i in range(pool_size):
            if i == correct_pos:
                candidates_data.append(description)
            else:
                if distractor_figs:
                    d_fig = distractor_figs.pop(0)
                    d_desc = agent_a_describe(client, title, int(d_fig["figure_number"]))
                    candidates_data.append(d_desc or {"viewpoint": "unknown", "visible_components": [], "spatial_relations": "", "distinguishing_features": ""})
                    time.sleep(0.3)
                else:
                    candidates_data.append({"viewpoint": "unknown", "visible_components": [], "spatial_relations": "", "distinguishing_features": ""})

        # Agent B: select the correct candidate
        chosen = agent_b_select(client, description, candidates_data)
        time.sleep(0.5)

        ok = chosen == correct_pos
        correct += int(ok)
        total += 1

        result = {
            "patent_id": pid,
            "patent_title": title,
            "target_fig": fig_num,
            "target_vp": target.get("vp", ""),
            "description": description,
            "correct_pos": correct_pos,
            "chosen": chosen,
            "correct": ok,
        }
        results.append(result)
        print(f"[{pid}] {title[:40]} | vp: {target.get('vp','?')[:20]} | {'✓' if ok else '✗'} | running: {correct}/{total} ({correct/total:.1%})")

    print(f"\n=== AGENT-AGENT RESULTS ===")
    print(f"Same-model ({get_model()}) → same-model")
    print(f"Accuracy: {correct}/{total} = {correct/total:.1%}")
    print(f"Chance (pool={pool_size}): {1/pool_size:.1%}")

    output = {
        "model": get_model(),
        "pool_size": pool_size,
        "summary": {"acc": correct / max(total, 1), "n": total, "chance": 1/pool_size},
        "results": results,
    }
    Path(out_path).parent.mkdir(parents=True, exist_ok=True)
    with open(out_path, "w") as f:
        json.dump(output, f, indent=2)
    print(f"Results → {out_path}")


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--enriched", default="data/enriched/enriched_2022.parquet")
    parser.add_argument("--n",        type=int, default=18)
    parser.add_argument("--pool",     type=int, default=4)
    parser.add_argument("--out",      default="results/agent_agent_results.json")
    args = parser.parse_args()
    run(args.enriched, args.n, args.out, args.pool)


if __name__ == "__main__":
    main()