From 2b414d13b7fe27dafa06eb44732fca833476a759 Mon Sep 17 00:00:00 2001
From: Macky <macky@localhost>
Date: Sun, 9 Aug 2026 12:23:15 +0700
Subject: [PATCH] fix(analyze): don't hard-fail when LLM generates <15 personas
 (retry + accept)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Live test found: deepseek returned 14/15 personas -> whole analyze 500'd, breaking the
'กดสร้าง -> auto-analyze' flow. Now generate() retries up to 3x with a nudge, and accepts
a short result (>=8 personas) instead of crashing — admin can top up the rest with
'สร้างบุคคลต้นแบบเพิ่มเติม'. Also default tolerance/recontact on generated personas.
All backend suites pass.
---
 backend/app/services/persona_generator.py | 30 ++++++++++++++++++-----
 1 file changed, 24 insertions(+), 6 deletions(-)

diff --git a/backend/app/services/persona_generator.py b/backend/app/services/persona_generator.py
index b9d8e95..e24ce98 100644
--- a/backend/app/services/persona_generator.py
+++ b/backend/app/services/persona_generator.py
@@ -31,10 +31,23 @@ class PersonaGenerator:
             f"Sales Kit:\n{kit_json}\n\n"
             f"Generate exactly 15 personas (5 per tier A/B/C) as JSON."
         )
-        result = self.llm.complete_json(
-            PERSONA_SYSTEM, user_prompt, temperature=0.8, max_tokens=14000
-        )
-        personas = result.get("personas") or []
+        # Real LLMs sometimes return fewer than 15 (truncation / merge). Retry up to 2 extra
+        # times with a nudge; the body below already ACCEPTS short results (>= 8) instead of
+        # hard-failing, so these retries are just best-effort to reach a fuller set.
+        attempt = 0
+        while True:
+            attempt += 1
+            prompt = user_prompt + (
+                ""
+                if attempt == 1
+                else "\n\n(Note: you left some personas out — please output all 15, one JSON object per persona, no extra prose.)"
+            )
+            result = self.llm.complete_json(
+                PERSONA_SYSTEM, prompt, temperature=0.8, max_tokens=14000
+            )
+            personas = result.get("personas") or []
+            if (isinstance(personas, list) and len(personas) >= 15) or attempt >= 3:
+                break
         if not isinstance(personas, list) or not personas:
             raise ValueError("persona generator returned no personas")
 
@@ -57,6 +70,8 @@ class PersonaGenerator:
             p.setdefault("pains", [])
             p.setdefault("negotiation_levers", [])
             p.setdefault("objections", [])
+            p.setdefault("tolerance", 3)
+            p.setdefault("recontact", False)
             normalized.append(p)
 
         # Wrap tier-C: ensure at least one wrong_text persona
@@ -69,6 +84,9 @@ class PersonaGenerator:
                     p["special"] = "wrong_text"
                     break
 
-        if len(normalized) < 15:
-            raise ValueError(f"expected 15 personas, generated {len(normalized)}")
+        if len(normalized) < 8:
+            raise ValueError(f"expected ~15 personas, generated only {len(normalized)}")
+        # NOTE: if we're short of 15 (real LLMs occasionally return 14/13), we ACCEPT what we
+        # got rather than crashing the whole analyze — the caller/UI can top up with
+        # "create more personas". A retry loop lives in generate().
         return normalized