Spaces:
Sleeping
Sleeping
Commit ·
fdbd75a
1
Parent(s): f169fdd
Fix JSON Crash: Lower Temp to 0.3 + Enhance Repair Logic + Add Error Fallback
Browse files- services/analyzer.py +36 -23
services/analyzer.py
CHANGED
|
@@ -41,19 +41,28 @@ def clean_json_string(text: str) -> str:
|
|
| 41 |
def repair_json(json_str: str) -> str:
|
| 42 |
"""Attempt to repair truncated JSON by closing open brackets/braces."""
|
| 43 |
json_str = json_str.strip()
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 52 |
return json_str
|
| 53 |
|
| 54 |
def analyze_seo_content(content: str) -> dict:
|
| 55 |
"""
|
| 56 |
-
Analyzes content using LOCAL Qwen-
|
| 57 |
"""
|
| 58 |
if pipe is None:
|
| 59 |
return {"error": "Model failed to load on server startup. Check logs."}
|
|
@@ -75,22 +84,22 @@ def analyze_seo_content(content: str) -> dict:
|
|
| 75 |
"content": f"""Analyze topic: "{content[:2000]}"
|
| 76 |
|
| 77 |
REQUIREMENTS:
|
| 78 |
-
- **Core Keywords**: 30 high-value keywords
|
| 79 |
- **Related Phrases**: 10 variations.
|
| 80 |
- **Viral Hashtags**: 8 tags.
|
| 81 |
-
- **Content Titles**: 3
|
| 82 |
-
- **Strategy Tips**: 3
|
| 83 |
-
- **Target Audience**:
|
| 84 |
|
| 85 |
JSON OUTPUT STRUCTURE:
|
| 86 |
{{
|
| 87 |
"core_keywords": [
|
| 88 |
{{"keyword": "...", "search_volume": "...", "competition": "...", "relevance": 90}}, ... (Total 30 items)
|
| 89 |
],
|
| 90 |
-
"related_phrases": ["..."
|
| 91 |
-
"viral_hashtags": [{{"tag": "#...", "post_count": "..."}}
|
| 92 |
-
"content_titles": ["..."
|
| 93 |
-
"strategy_tips": ["..."
|
| 94 |
"target_audience": "..."
|
| 95 |
}}
|
| 96 |
"""
|
|
@@ -101,16 +110,13 @@ def analyze_seo_content(content: str) -> dict:
|
|
| 101 |
# Qwen/HuggingFace pipeline handles chat templates automatically
|
| 102 |
outputs = pipe(
|
| 103 |
messages,
|
| 104 |
-
max_new_tokens=1500,
|
| 105 |
do_sample=True,
|
| 106 |
-
temperature=0.
|
| 107 |
top_p=0.9
|
| 108 |
)
|
| 109 |
|
| 110 |
-
# Extract generated text
|
| 111 |
content_text = outputs[0]["generated_text"][-1]["content"]
|
| 112 |
-
|
| 113 |
-
# Clean and Parse
|
| 114 |
cleaned_json = clean_json_string(content_text)
|
| 115 |
|
| 116 |
try:
|
|
@@ -118,7 +124,14 @@ def analyze_seo_content(content: str) -> dict:
|
|
| 118 |
except json.JSONDecodeError:
|
| 119 |
print("⚠️ JSON Parse Error. Attempting repair...")
|
| 120 |
cleaned_json = repair_json(cleaned_json)
|
| 121 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 122 |
|
| 123 |
return data
|
| 124 |
|
|
|
|
| 41 |
def repair_json(json_str: str) -> str:
|
| 42 |
"""Attempt to repair truncated JSON by closing open brackets/braces."""
|
| 43 |
json_str = json_str.strip()
|
| 44 |
+
|
| 45 |
+
# Remove potentially invalid trailing content that isn't a closer
|
| 46 |
+
# (e.g. if it ends with "keyword": "val", -> we need to remove comma)
|
| 47 |
+
json_str = json_str.rstrip(", ")
|
| 48 |
+
|
| 49 |
+
# Balance brackets
|
| 50 |
+
open_braces = json_str.count("{")
|
| 51 |
+
close_braces = json_str.count("}")
|
| 52 |
+
open_brackets = json_str.count("[")
|
| 53 |
+
close_brackets = json_str.count("]")
|
| 54 |
+
|
| 55 |
+
if open_braces > close_braces:
|
| 56 |
+
json_str += "}" * (open_braces - close_braces)
|
| 57 |
+
|
| 58 |
+
if open_brackets > close_brackets:
|
| 59 |
+
json_str += "]" * (open_brackets - close_brackets)
|
| 60 |
+
|
| 61 |
return json_str
|
| 62 |
|
| 63 |
def analyze_seo_content(content: str) -> dict:
|
| 64 |
"""
|
| 65 |
+
Analyzes content using LOCAL Qwen-0.5B.
|
| 66 |
"""
|
| 67 |
if pipe is None:
|
| 68 |
return {"error": "Model failed to load on server startup. Check logs."}
|
|
|
|
| 84 |
"content": f"""Analyze topic: "{content[:2000]}"
|
| 85 |
|
| 86 |
REQUIREMENTS:
|
| 87 |
+
- **Core Keywords**: 30 high-value keywords.
|
| 88 |
- **Related Phrases**: 10 variations.
|
| 89 |
- **Viral Hashtags**: 8 tags.
|
| 90 |
+
- **Content Titles**: 3 titles.
|
| 91 |
+
- **Strategy Tips**: 3 tips.
|
| 92 |
+
- **Target Audience**: Profile.
|
| 93 |
|
| 94 |
JSON OUTPUT STRUCTURE:
|
| 95 |
{{
|
| 96 |
"core_keywords": [
|
| 97 |
{{"keyword": "...", "search_volume": "...", "competition": "...", "relevance": 90}}, ... (Total 30 items)
|
| 98 |
],
|
| 99 |
+
"related_phrases": ["..."],
|
| 100 |
+
"viral_hashtags": [{{"tag": "#...", "post_count": "..."}}],
|
| 101 |
+
"content_titles": ["..."],
|
| 102 |
+
"strategy_tips": ["..."],
|
| 103 |
"target_audience": "..."
|
| 104 |
}}
|
| 105 |
"""
|
|
|
|
| 110 |
# Qwen/HuggingFace pipeline handles chat templates automatically
|
| 111 |
outputs = pipe(
|
| 112 |
messages,
|
| 113 |
+
max_new_tokens=1500,
|
| 114 |
do_sample=True,
|
| 115 |
+
temperature=0.3, # Low temp for strict JSON
|
| 116 |
top_p=0.9
|
| 117 |
)
|
| 118 |
|
|
|
|
| 119 |
content_text = outputs[0]["generated_text"][-1]["content"]
|
|
|
|
|
|
|
| 120 |
cleaned_json = clean_json_string(content_text)
|
| 121 |
|
| 122 |
try:
|
|
|
|
| 124 |
except json.JSONDecodeError:
|
| 125 |
print("⚠️ JSON Parse Error. Attempting repair...")
|
| 126 |
cleaned_json = repair_json(cleaned_json)
|
| 127 |
+
try:
|
| 128 |
+
data = json.loads(cleaned_json)
|
| 129 |
+
except json.JSONDecodeError as final_err:
|
| 130 |
+
print(f"❌ Repair Failed: {final_err}")
|
| 131 |
+
return {
|
| 132 |
+
"error": "Optimization Failed. The model generated invalid JSON.",
|
| 133 |
+
"raw_output": content_text[:500] + "..." # Debug info
|
| 134 |
+
}
|
| 135 |
|
| 136 |
return data
|
| 137 |
|