ihtesham0345 commited on
Commit
e1cb6ad
·
1 Parent(s): b0fb0e7

fix: Add normalize_response to handle model output format mismatches

Browse files

- Model returns strings where objects expected (e.g., video_titles: ['str'] → [{'title':'str','expected_ctr':'Medium'}])
- Model returns objects where strings expected (e.g., thumbnail_ideas: [{url, caption}] → ['caption text'])
- Normalizes all field types: keyword_list, hashtag_list, caption_list, pin_list, object_list, string_list
- Handles nested structures: hashtag_sets, content_ideas, post_drafts, tweet_threads

Files changed (1) hide show
  1. services/utils.py +122 -0
services/utils.py CHANGED
@@ -92,6 +92,126 @@ def repair_json_v2(json_str: str) -> str:
92
 
93
  return trimmed
94
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
95
  def run_analysis(messages, defaults=None, temperature=0.3, max_new_tokens=2000):
96
  from services.model_loader import generate_text
97
  raw = generate_text(messages, temperature, max_new_tokens)
@@ -106,6 +226,8 @@ def run_analysis(messages, defaults=None, temperature=0.3, max_new_tokens=2000):
106
  base["error"] = f"JSON parse failed: {err.get('error', 'unknown')}"
107
  return base
108
 
 
 
109
  if defaults:
110
  return validate_and_fill_data_defaults(data, defaults)
111
  return data
 
92
 
93
  return trimmed
94
 
95
+ def normalize_field(items, expected_type):
96
+ """Convert model output to match Pydantic expectations."""
97
+ if not isinstance(items, list):
98
+ return items
99
+
100
+ if expected_type == "object_list":
101
+ # Items should be [{key: val, ...}, ...] but model gave ["str1", "str2"]
102
+ if items and isinstance(items[0], str):
103
+ return [{"title": s, "expected_ctr": "Medium"} for s in items]
104
+ return items
105
+
106
+ if expected_type == "string_list":
107
+ # Items should be ["str1", "str2"] but model gave [{...}, {...}]
108
+ if items and isinstance(items[0], dict):
109
+ extracted = []
110
+ for obj in items:
111
+ val = obj.get("caption") or obj.get("description") or obj.get("title") or obj.get("name") or obj.get("url") or str(obj)
112
+ extracted.append(val)
113
+ return extracted
114
+ return items
115
+
116
+ if expected_type == "keyword_list":
117
+ # Items should be [{keyword, search_volume, competition, relevance}]
118
+ if items and isinstance(items[0], str):
119
+ return [{"keyword": s, "search_volume": "Medium", "competition": "Medium", "relevance": 80} for s in items]
120
+ return items
121
+
122
+ if expected_type == "hashtag_list":
123
+ # Items should be [{tag, post_count}] but model gave ["#tag1", "#tag2"]
124
+ if items and isinstance(items[0], str):
125
+ return [{"tag": s, "post_count": "N/A"} for s in items]
126
+ return items
127
+
128
+ if expected_type == "caption_list":
129
+ if items and isinstance(items[0], str):
130
+ return [{"caption": s, "tone": "engaging"} for s in items]
131
+ return items
132
+
133
+ if expected_type == "pin_list":
134
+ if items and isinstance(items[0], str):
135
+ return [{"title": s, "description": s, "keyword_focus": s} for s in items]
136
+ return items
137
+
138
+ return items
139
+
140
+ def normalize_response(data: dict) -> dict:
141
+ """Post-process model output to match expected Pydantic schemas."""
142
+ # YouTube: video_titles expects list of {title, expected_ctr}
143
+ if "video_titles" in data:
144
+ data["video_titles"] = normalize_field(data["video_titles"], "object_list")
145
+ if "thumbnail_ideas" in data:
146
+ data["thumbnail_ideas"] = normalize_field(data["thumbnail_ideas"], "string_list")
147
+
148
+ # Instagram
149
+ if "captions" in data:
150
+ data["captions"] = normalize_field(data["captions"], "caption_list")
151
+ if "hashtag_sets" in data and isinstance(data["hashtag_sets"], dict):
152
+ for tier in ["small", "medium", "large"]:
153
+ if tier in data["hashtag_sets"] and isinstance(data["hashtag_sets"][tier], list):
154
+ if data["hashtag_sets"][tier] and isinstance(data["hashtag_sets"][tier][0], dict):
155
+ data["hashtag_sets"][tier] = [obj.get("tag", obj.get("name", str(obj))) for obj in data["hashtag_sets"][tier]]
156
+ if "content_ideas" in data and isinstance(data["content_ideas"], dict):
157
+ for key in ["reels", "carousels", "stories"]:
158
+ if key in data["content_ideas"]:
159
+ data["content_ideas"][key] = normalize_field(data["content_ideas"][key], "string_list")
160
+
161
+ # Generic SEO
162
+ if "core_keywords" in data:
163
+ data["core_keywords"] = normalize_field(data["core_keywords"], "keyword_list")
164
+ if "viral_hashtags" in data:
165
+ data["viral_hashtags"] = normalize_field(data["viral_hashtags"], "hashtag_list")
166
+
167
+ # Pinterest
168
+ if "pin_ideas" in data:
169
+ data["pin_ideas"] = normalize_field(data["pin_ideas"], "pin_list")
170
+ if "seo_keywords" in data:
171
+ data["seo_keywords"] = normalize_field(data["seo_keywords"], "string_list")
172
+
173
+ # Generic string lists that might have gotten objects
174
+ for field in ["related_phrases", "strategy_tips", "content_titles",
175
+ "tags", "engagement_strategies", "growth_strategies",
176
+ "article_topics", "thought_leadership_angles",
177
+ "engagement_hooks", "ad_copy_suggestions", "page_growth_tips",
178
+ "viral_hooks", "engagement_tactics",
179
+ "trending_angles", "viral_strategies",
180
+ "board_organization", "traffic_strategies",
181
+ "hashtags"]:
182
+ if field in data:
183
+ data[field] = normalize_field(data[field], "string_list")
184
+
185
+ # LinkedIn post_drafts: should be [{headline, body, hook}]
186
+ if "post_drafts" in data:
187
+ if data["post_drafts"] and isinstance(data["post_drafts"][0], str):
188
+ data["post_drafts"] = [{"headline": s, "body": s, "hook": s} for s in data["post_drafts"]]
189
+
190
+ # Post ideas: should be [{type, content}]
191
+ if "post_ideas" in data:
192
+ if data["post_ideas"] and isinstance(data["post_ideas"][0], str):
193
+ data["post_ideas"] = [{"type": "text", "content": s} for s in data["post_ideas"]]
194
+
195
+ # Tweet threads: should be [{tweets: [str], theme: str}]
196
+ if "tweet_threads" in data:
197
+ if data["tweet_threads"] and isinstance(data["tweet_threads"][0], str):
198
+ data["tweet_threads"] = [{"tweets": [data["tweet_threads"][0]], "theme": "topic"}]
199
+
200
+ # Video concepts: should be [{hook, script_snippet, sound_suggestion}]
201
+ if "video_concepts" in data:
202
+ if data["video_concepts"] and isinstance(data["video_concepts"][0], str):
203
+ data["video_concepts"] = [{"hook": s, "script_snippet": s, "sound_suggestion": "Trending"} for s in data["video_concepts"]]
204
+
205
+ # Ensure descriptions are strings, not objects
206
+ if "description_template" in data and isinstance(data["description_template"], dict):
207
+ data["description_template"] = str(data["description_template"])
208
+
209
+ # Target audience should be string
210
+ if "target_audience" in data and isinstance(data["target_audience"], dict):
211
+ data["target_audience"] = str(data["target_audience"])
212
+
213
+ return data
214
+
215
  def run_analysis(messages, defaults=None, temperature=0.3, max_new_tokens=2000):
216
  from services.model_loader import generate_text
217
  raw = generate_text(messages, temperature, max_new_tokens)
 
226
  base["error"] = f"JSON parse failed: {err.get('error', 'unknown')}"
227
  return base
228
 
229
+ data = normalize_response(data)
230
+
231
  if defaults:
232
  return validate_and_fill_data_defaults(data, defaults)
233
  return data