File size: 10,079 Bytes
15c3265
 
 
 
 
 
 
9b7e0a7
15c3265
 
 
9b7e0a7
 
15c3265
 
 
 
 
 
 
9b7e0a7
15c3265
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9b7e0a7
15c3265
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9b7e0a7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
15c3265
9b7e0a7
 
 
 
 
 
 
15c3265
 
 
 
 
 
 
 
 
 
 
9b7e0a7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
15c3265
9b7e0a7
 
 
 
 
 
 
15c3265
 
 
 
 
 
 
 
 
9b7e0a7
15c3265
9b7e0a7
 
15c3265
 
 
 
 
 
9b7e0a7
15c3265
 
 
 
 
 
 
 
 
9b7e0a7
15c3265
9b7e0a7
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
"""Response composition with fact-first persona-biased fusion."""

from __future__ import annotations

from dataclasses import dataclass
import json
from pathlib import Path
from typing import Any, Dict, List, Optional, Sequence

from ..config import RuntimeConfig
from ..retrieval.retriever import tri_retrieve
from ..utils.retry import retry_call
from .token_usage import new_token_usage, record_token_usage


@dataclass
class FusionResult:
    answer: str
    draft: str
    evidence: Dict[str, object]
    token_usage: Dict[str, Any]


def _load_persona_profile(cfg: RuntimeConfig, character: Optional[str]) -> Dict:
    if not character:
        return {}
    safe = character.replace("/", "_")
    profile_path = Path(cfg.data_processed_dir) / f"persona_{safe}.json"
    if not profile_path.exists():
        return {}
    try:
        with open(profile_path, "r", encoding="utf-8") as f:
            return json.load(f)
    except Exception:
        return {}


def _format_lane_rows(rows: List[Dict], lane_tag: str, max_rows: int = 6) -> str:
    if not rows:
        return "(none)"
    lines: List[str] = []
    for i, row in enumerate(rows[:max_rows], start=1):
        text = str(row.get("text", "")).strip().replace("\n", " ")
        score = float(row.get("score", 0.0))
        extra = ""
        if lane_tag == "P":
            speaker = row.get("speaker")
            extra = f" speaker={speaker}" if speaker else ""
        if lane_tag == "W":
            t = row.get("type")
            ent = row.get("entity")
            parts = []
            if t:
                parts.append(f"type={t}")
            if ent:
                parts.append(f"entity={ent}")
            extra = (" " + " ".join(parts)) if parts else ""
        lines.append(f"[{lane_tag}{i}] score={score:.3f}{extra} text={text}")
    return "\n".join(lines)


def _build_fusion_prompts(
    query: str,
    evidence: Dict[str, object],
    character: Optional[str],
    persona_profile: Dict,
) -> tuple[str, str]:
    facts = evidence.get("facts", [])
    persona = evidence.get("persona", [])
    worldview = evidence.get("worldview", [])
    query_plan = evidence.get("query_plan", {})

    style_markers = persona_profile.get("style_markers", [])
    speaking_rules = persona_profile.get("speaking_rules", [])
    values = persona_profile.get("values", [])

    system_prompt = (
        "You are a retrieval-grounded fiction dialogue assistant.\n"
        "Priority policy:\n"
        "1) Facts evidence is highest priority.\n"
        "2) Worldview evidence is second priority.\n"
        "3) Persona style is third priority.\n"
        "If style conflicts with facts/worldview, keep facts/worldview.\n"
        "Do not invent unsupported facts.\n"
        "When possible, cite evidence IDs like [F1], [W2], [P1]."
    )

    user_prompt = (
        f"User query:\n{query}\n\n"
        f"Character:\n{character or 'Unknown'}\n\n"
        f"Query decomposition:\n{json.dumps(query_plan, ensure_ascii=False)}\n\n"
        f"Character style markers:\n{', '.join(style_markers) if style_markers else '(none)'}\n"
        f"Character speaking rules:\n{'; '.join(speaking_rules) if speaking_rules else '(none)'}\n"
        f"Character values:\n{', '.join(values) if values else '(none)'}\n\n"
        f"Facts evidence:\n{_format_lane_rows(facts, 'F')}\n\n"
        f"Worldview evidence:\n{_format_lane_rows(worldview, 'W')}\n\n"
        f"Persona evidence:\n{_format_lane_rows(persona, 'P')}\n\n"
        "Generate a concise in-character answer with factual consistency."
    )
    return system_prompt, user_prompt


def _build_style_correction_prompts(
    draft: str,
    character: Optional[str],
    persona_profile: Dict,
) -> tuple[str, str]:
    style_markers = persona_profile.get("style_markers", [])
    speaking_rules = persona_profile.get("speaking_rules", [])
    examples = persona_profile.get("examples", [])

    system_prompt = (
        "You are a style editor for fictional roleplay.\n"
        "Rewrite only style/tone/wording to match the target character.\n"
        "Do NOT change factual content or claims.\n"
        "Do NOT add new facts."
    )
    user_prompt = (
        f"Character: {character or 'Unknown'}\n"
        f"Style markers: {', '.join(style_markers) if style_markers else '(none)'}\n"
        f"Speaking rules: {'; '.join(speaking_rules) if speaking_rules else '(none)'}\n"
        f"Example lines: {' | '.join(examples[:3]) if examples else '(none)'}\n\n"
        f"Draft answer:\n{draft}\n\n"
        "Return revised answer only."
    )
    return system_prompt, user_prompt


def _heuristic_fallback_answer(query: str, evidence: Dict[str, object], character: Optional[str]) -> str:
    facts = evidence.get("facts", [])[:3]
    worldview = evidence.get("worldview", [])[:2]
    persona = evidence.get("persona", [])[:2]

    lines = [f"Q: {query}", ""]
    if character:
        lines.append(f"Character: {character}")
    lines.append("Fact-grounded points:")
    if facts:
        for row in facts:
            lines.append(f"- {row.get('text', '')}")
    else:
        lines.append("- (no strong fact evidence found)")
    lines.append("World consistency notes:")
    if worldview:
        for row in worldview:
            lines.append(f"- {row.get('text', '')}")
    else:
        lines.append("- (no worldview constraints retrieved)")
    lines.append("Persona cues:")
    if persona:
        for row in persona:
            lines.append(f"- {row.get('text', '')}")
    else:
        lines.append("- (no persona cue retrieved)")
    return "\n".join(lines)


def compose_response(
    query: str,
    evidence: Dict[str, object],
    cfg: RuntimeConfig,
    character: Optional[str] = None,
    style_correct: bool = False,
    token_usage: Optional[Dict[str, Any]] = None,
) -> str:
    profile = _load_persona_profile(cfg, character)
    if not cfg.llm_api_key:
        return _heuristic_fallback_answer(query=query, evidence=evidence, character=character)

    try:
        from openai import OpenAI  # type: ignore
    except Exception as e:
        raise ImportError(
            "openai package is required for LLM composition. Install dependencies first."
        ) from e

    client = OpenAI(base_url=cfg.llm_base_url, api_key=cfg.llm_api_key)
    sys_prompt, user_prompt = _build_fusion_prompts(
        query=query,
        evidence=evidence,
        character=character,
        persona_profile=profile,
    )
    def _draft_call():
        return client.chat.completions.create(
            model=cfg.llm_model,
            messages=[
                {"role": "system", "content": sys_prompt},
                {"role": "user", "content": user_prompt},
            ],
            temperature=0.2,
            max_tokens=700,
        )

    def _on_draft_retry(attempt: int, err: Exception, delay: float) -> None:
        print(
            f"[compose][retry] stage=draft attempt={attempt + 1}/{max(1, int(cfg.api_retry_attempts))} sleep={delay:.1f}s err={err}",
            flush=True,
        )

    draft_resp = retry_call(
        _draft_call,
        max_attempts=max(1, int(cfg.api_retry_attempts)),
        base_delay_sec=float(cfg.api_retry_base_delay_sec),
        max_delay_sec=float(cfg.api_retry_max_delay_sec),
        jitter_sec=float(cfg.api_retry_jitter_sec),
        on_retry=_on_draft_retry,
    )
    if token_usage is not None:
        record_token_usage(
            token_usage,
            response=draft_resp,
            stage="compose_draft",
            model=cfg.llm_model,
        )
    draft = (draft_resp.choices[0].message.content or "").strip()
    if not draft:
        return _heuristic_fallback_answer(query=query, evidence=evidence, character=character)
    if not style_correct:
        return draft

    style_sys, style_user = _build_style_correction_prompts(
        draft=draft,
        character=character,
        persona_profile=profile,
    )
    def _style_call():
        return client.chat.completions.create(
            model=cfg.llm_model,
            messages=[
                {"role": "system", "content": style_sys},
                {"role": "user", "content": style_user},
            ],
            temperature=0.1,
            max_tokens=700,
        )

    def _on_style_retry(attempt: int, err: Exception, delay: float) -> None:
        print(
            f"[compose][retry] stage=style_rewrite attempt={attempt + 1}/{max(1, int(cfg.api_retry_attempts))} sleep={delay:.1f}s err={err}",
            flush=True,
        )

    style_resp = retry_call(
        _style_call,
        max_attempts=max(1, int(cfg.api_retry_attempts)),
        base_delay_sec=float(cfg.api_retry_base_delay_sec),
        max_delay_sec=float(cfg.api_retry_max_delay_sec),
        jitter_sec=float(cfg.api_retry_jitter_sec),
        on_retry=_on_style_retry,
    )
    if token_usage is not None:
        record_token_usage(
            token_usage,
            response=style_resp,
            stage="style_rewrite",
            model=cfg.llm_model,
        )
    revised = (style_resp.choices[0].message.content or "").strip()
    return revised or draft


def run_tri_retrieve_and_compose(
    query: str,
    cfg: RuntimeConfig,
    character: Optional[str] = None,
    style_correct: bool = False,
    active_lanes: Optional[Sequence[str]] = None,
) -> FusionResult:
    token_usage: Dict[str, Any] = new_token_usage()
    evidence = tri_retrieve(query=query, cfg=cfg, character=character, active_lanes=active_lanes)
    draft = compose_response(
        query=query,
        evidence=evidence,
        cfg=cfg,
        character=character,
        style_correct=False,
        token_usage=token_usage,
    )
    final_answer = draft
    if style_correct:
        final_answer = compose_response(
            query=query,
            evidence=evidence,
            cfg=cfg,
            character=character,
            style_correct=True,
            token_usage=token_usage,
        )
    return FusionResult(answer=final_answer, draft=draft, evidence=evidence, token_usage=token_usage)