Recover from malformed structured model output

app.py

@@ -71,25 +71,25 @@ SectionLayout = Literal[
 
 class SectionPlan(BaseModel):
     id: str = Field(min_length=2)
-    title: str = Field(min_length=2)
-    purpose: str = Field(min_length=10)
+    title: str = Field(min_length=2, max_length=120)
+    purpose: str = Field(min_length=10, max_length=500)
     layout: SectionLayout
     source_ids: list[str] = Field(default_factory=list)
 
 
 class ReportOutline(BaseModel):
-    title: str = Field(min_length=2)
-    subtitle: str = Field(min_length=2)
-    research_type: str = Field(min_length=3)
+    title: str = Field(min_length=2, max_length=160)
+    subtitle: str = Field(min_length=2, max_length=240)
+    research_type: str = Field(min_length=3, max_length=80)
     executive_summary: list[str] = Field(min_length=1, max_length=6)
     section_plan: list[SectionPlan] = Field(min_length=3, max_length=12)
 
 
 class ReportSection(BaseModel):
     id: str = Field(min_length=2)
-    title: str = Field(min_length=2)
-    purpose: str = Field(min_length=10)
-    takeaway: str = Field(min_length=10)
+    title: str = Field(min_length=2, max_length=120)
+    purpose: str = Field(min_length=10, max_length=500)
+    takeaway: str = Field(min_length=10, max_length=800)
     layout: SectionLayout
     content: dict[str, Any]
     source_ids: list[str] = Field(default_factory=list)
@@ -125,9 +125,9 @@ class SectionsEnvelope(BaseModel):
 
 
 class ResearchReport(BaseModel):
-    title: str = Field(min_length=2)
-    subtitle: str = Field(min_length=2)
-    research_type: str = Field(min_length=3)
+    title: str = Field(min_length=2, max_length=160)
+    subtitle: str = Field(min_length=2, max_length=240)
+    research_type: str = Field(min_length=3, max_length=80)
     executive_summary: list[str] = Field(min_length=1, max_length=6)
     sections: list[ReportSection] = Field(min_length=3, max_length=12)
     references: list[str] = Field(default_factory=list)
@@ -362,12 +362,14 @@ def generate_text(system: str, prompt: str, max_tokens: int, structured: bool =
     return str(content).strip()
 
 
-def chat_json(system: str, prompt: str, max_tokens: int = 6000) -> dict[str, Any]:
-    content = generate_text(system, prompt, max_tokens)
+def parse_json_object(content: str) -> dict[str, Any] | None:
     candidates = [content]
     fenced = re.sub(r"^```(?:json)?\s*|\s*```$", "", content, flags=re.IGNORECASE)
     if fenced != content:
         candidates.append(fenced)
+    line_continuations_removed = re.sub(r"\\\r?\n", "", fenced)
+    if line_continuations_removed != fenced:
+        candidates.append(line_continuations_removed)
     match = re.search(r"\{[\s\S]*\}", content)
     if match:
         candidates.append(match.group(0))
@@ -378,7 +380,29 @@ def chat_json(system: str, prompt: str, max_tokens: int = 6000) -> dict[str, Any
                 return parsed
         except (json.JSONDecodeError, TypeError):
             continue
-    preview = content[:240].replace("\n", " ") if content else "<empty response>"
+    return None
+
+
+def chat_json(system: str, prompt: str, max_tokens: int = 6000) -> dict[str, Any]:
+    last_content = ""
+    for attempt in range(2):
+        retry_rules = ""
+        retry_prompt = prompt
+        if attempt:
+            retry_rules = (
+                " The previous response was malformed or degenerated. Regenerate from scratch as compact valid JSON."
+                " Never put literal or escaped line breaks in string values. Do not repeat whitespace or characters."
+                " Keep titles under 120 characters and every other string under 500 characters. Close every array and object."
+            )
+            retry_prompt = (
+                f"{prompt}\n\nThe previous response could not be parsed. Its beginning was: "
+                f"{last_content[:320]!r}\nRegenerate the requested object from scratch."
+            )
+        last_content = generate_text(system + retry_rules, retry_prompt, max_tokens)
+        parsed = parse_json_object(last_content)
+        if parsed is not None:
+            return parsed
+    preview = last_content[:240].replace("\n", " ") if last_content else "<empty response>"
     raise RuntimeError(
         f"{MODEL_ID} did not return a JSON object. Provider={HF_PROVIDER}. Response preview: {preview}"
     )

tests/test_app.py

@@ -1,8 +1,9 @@
 import unittest
+from unittest.mock import patch
 
 from pydantic import ValidationError
 
-from app import ResearchReport, deduplicate, normalize_comparison, prepare_evidence_inputs, render_presentation, render_section_content, render_value, safe_slug
+from app import ResearchReport, chat_json, deduplicate, normalize_comparison, parse_json_object, prepare_evidence_inputs, render_presentation, render_section_content, render_value, safe_slug
 
 
 def base_report(section):
@@ -30,6 +31,15 @@ class AppTests(unittest.TestCase):
         self.assertEqual([item["id"] for item in result[:2]], ["8", "9"])
         self.assertIn("selected_text_extract", result[0])
 
+    def test_parser_removes_invalid_string_line_continuation(self):
+        self.assertEqual(parse_json_object('{"title":"A\\\nB","value":1}'), {"title": "AB", "value": 1})
+
+    @patch("app.generate_text", side_effect=['{"title":"broken\\\n', '{"title":"valid"}'])
+    def test_json_generation_retries_malformed_response(self, mocked_generate):
+        self.assertEqual(chat_json("system", "prompt", 2048), {"title": "valid"})
+        self.assertEqual(mocked_generate.call_count, 2)
+        self.assertIn("Regenerate from scratch", mocked_generate.call_args.args[0])
+
     def test_render_is_downloadable_document(self):
         section = {"title": "Scope", "purpose": "Define the technical decision boundary.", "takeaway": "The boundary determines the useful evidence.", "layout": "text", "content": {"points": ["A meaningful point", "Another meaningful point"]}, "source_ids": []}
         output = render_presentation(base_report(section), [], "2026-01-01")