Sync from GitHub · 2026-09-09 19:46 UTC

routes/bff.py

@@ -50,8 +50,15 @@ def _get_bff_data(user_id: str, business_id: str) -> dict:
 
 
 def _call_groq_json(system_prompt: str, user_prompt: str) -> dict:
-    """Thin wrapper — delegates to the key-rotating Groq client utility."""
-    return call_groq_json(system_prompt, user_prompt, model=HEAVY_MODEL, max_tokens=1500)
+    """
+    Thin wrapper — delegates to the key-rotating Groq client utility.
+
+    max_tokens is generous (3000) because HEAVY_MODEL is a reasoning model
+    (openai/gpt-oss-120b) that spends part of its budget on internal
+    reasoning before writing the final JSON — a tight budget here truncates
+    the JSON output and fails to parse.
+    """
+    return call_groq_json(system_prompt, user_prompt, model=HEAVY_MODEL, max_tokens=3000)
 
 
 # ─── Route: Generate follow-up questions ─────────────────────────────────────
@@ -270,19 +277,23 @@ each containing a list of 2-4 concise bullet points:
     }
 
     call_plan = []
+    skipped_keys: set = set()
     for key, prompt_fn in all_outputs.items():
         existing = biz_ref_check.collection("profile_output").document(key).get()
         if existing.exists:
             print(f"[bff] Skipping {key} — already exists in Firestore")
+            skipped_keys.add(key)
         else:
             call_plan.append((key, prompt_fn))
 
     results: dict = {}
     errors: list = []
 
-    # Sequential calls with 15s gap to stay within 12,000 TPM free-tier limit.
-    # With 1,500 max_tokens per call: ~2,500 tokens/call × 4 = ~10,000 total.
-    # 15s gap × 200 tokens/sec refill = 3,000 tokens recovered between calls.
+    # Sequential calls with 15s gap to stay within the free-tier TPM limit.
+    # max_tokens=3000 per call leaves headroom for HEAVY_MODEL's internal
+    # reasoning tokens on top of the JSON output (see _call_groq_json).
+    # 15s gap × 200 tokens/sec refill recovers ~3,000 tokens between calls —
+    # if that's not enough, the 5-key rotation pool absorbs the rest via 429s.
     for i, (key, prompt_fn) in enumerate(call_plan):
         if i > 0:
             time.sleep(15)  # 15s gap: ~3,000 tokens recovered at 200 tok/s
@@ -314,10 +325,22 @@ each containing a list of 2-4 concise bullet points:
             {**data, "generatedAt": SERVER_TIMESTAMP}
         )
 
-    # Mark BFF as complete
-    biz_ref.update({"bffCompleted": True})
+    # Only mark BFF as complete once EVERY output actually exists — either
+    # generated just now or already present from a prior partial run. Marking
+    # this true when an output is still missing is what previously let a
+    # business look "complete" with an empty/partial profile (see
+    # project_multibusiness_gate memory — the manual-Console-flip incident
+    # had the same shape of bug).
+    present_keys = skipped_keys | set(results.keys())
+    all_present  = present_keys == set(all_outputs.keys())
+    if all_present:
+        biz_ref.update({"bffCompleted": True})
+    else:
+        missing = set(all_outputs.keys()) - present_keys
+        print(f"[bff] NOT marking bffCompleted — missing outputs: {missing}")
 
     return {
-        "success": True,
+        "success": all_present,
         "outputs": list(results.keys()),
+        "missing": sorted(set(all_outputs.keys()) - present_keys),
     }

utils/groq_client.py

@@ -174,7 +174,17 @@ def call_groq_json(
     """Convenience wrapper: calls Groq in JSON mode and returns parsed dict."""
     raw = call_groq(system_prompt, user_prompt, model=model,
                     json_mode=True, max_tokens=max_tokens)
-    return json.loads(raw)
+    if not raw:
+        # Reasoning models (e.g. openai/gpt-oss-*) can spend the token budget on
+        # internal reasoning and return empty content if max_tokens is too tight.
+        raise RuntimeError(
+            "Groq returned empty content — likely truncated by max_tokens "
+            f"(model={model}, max_tokens={max_tokens})"
+        )
+    try:
+        return json.loads(raw)
+    except json.JSONDecodeError as e:
+        raise RuntimeError(f"Groq returned invalid JSON (likely truncated): {e}") from e
 
 
 def call_groq_stream(