Spaces:
Paused
Paused
Tescik 2
Browse files- brain/orchestrator.py +13 -11
brain/orchestrator.py
CHANGED
|
@@ -57,26 +57,25 @@ def _signoff_bubbles(ctx: dict):
|
|
| 57 |
]
|
| 58 |
|
| 59 |
|
| 60 |
-
# 🔴 FIX: speaker tagging added here
|
| 61 |
def _history_to_messages(ctx: dict):
|
| 62 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 63 |
|
|
|
|
| 64 |
media = ctx.get("media", {})
|
| 65 |
|
| 66 |
for h in ctx.get("history", []):
|
| 67 |
role = h.get("role", "user")
|
| 68 |
text = h.get("text", "")
|
| 69 |
|
| 70 |
-
#
|
| 71 |
-
|
| 72 |
-
content = f"[ARIA]: {text}"
|
| 73 |
-
else:
|
| 74 |
-
content = f"[USER]: {text}"
|
| 75 |
|
| 76 |
-
# multimodal handling
|
| 77 |
-
parts = [{"type": "text", "text": content}]
|
| 78 |
m = media.get(h.get("id"))
|
| 79 |
-
|
| 80 |
if m and m.get("kind") == "image" and m.get("url"):
|
| 81 |
parts.append({"type": "image_url", "image_url": {"url": m["url"]}})
|
| 82 |
|
|
@@ -86,6 +85,8 @@ def _history_to_messages(ctx: dict):
|
|
| 86 |
|
| 87 |
|
| 88 |
def respond(context_json: str) -> str:
|
|
|
|
|
|
|
| 89 |
try:
|
| 90 |
ctx = json.loads(context_json) if isinstance(context_json, str) else (context_json or {})
|
| 91 |
except Exception:
|
|
@@ -96,9 +97,10 @@ def respond(context_json: str) -> str:
|
|
| 96 |
|
| 97 |
messages = _history_to_messages(ctx)
|
| 98 |
|
| 99 |
-
#
|
| 100 |
messages.append({"role": "user", "content": build_user_frame(ctx)})
|
| 101 |
|
|
|
|
| 102 |
lang = _lang(ctx)
|
| 103 |
messages.append({
|
| 104 |
"role": "user",
|
|
|
|
| 57 |
]
|
| 58 |
|
| 59 |
|
|
|
|
| 60 |
def _history_to_messages(ctx: dict):
|
| 61 |
+
"""
|
| 62 |
+
IMPORTANT FIX:
|
| 63 |
+
- NO speaker tagging in text
|
| 64 |
+
- role metadata is the ONLY source of truth
|
| 65 |
+
- prevents HF chat template corruption + response dropout
|
| 66 |
+
"""
|
| 67 |
|
| 68 |
+
msgs = [{"role": "system", "content": SYSTEM_PROMPT}]
|
| 69 |
media = ctx.get("media", {})
|
| 70 |
|
| 71 |
for h in ctx.get("history", []):
|
| 72 |
role = h.get("role", "user")
|
| 73 |
text = h.get("text", "")
|
| 74 |
|
| 75 |
+
# keep raw text ONLY
|
| 76 |
+
parts = [{"type": "text", "text": text}]
|
|
|
|
|
|
|
|
|
|
| 77 |
|
|
|
|
|
|
|
| 78 |
m = media.get(h.get("id"))
|
|
|
|
| 79 |
if m and m.get("kind") == "image" and m.get("url"):
|
| 80 |
parts.append({"type": "image_url", "image_url": {"url": m["url"]}})
|
| 81 |
|
|
|
|
| 85 |
|
| 86 |
|
| 87 |
def respond(context_json: str) -> str:
|
| 88 |
+
"""Entry called by the Gradio API (api_name='respond')."""
|
| 89 |
+
|
| 90 |
try:
|
| 91 |
ctx = json.loads(context_json) if isinstance(context_json, str) else (context_json or {})
|
| 92 |
except Exception:
|
|
|
|
| 97 |
|
| 98 |
messages = _history_to_messages(ctx)
|
| 99 |
|
| 100 |
+
# frame that tells model what to do this turn
|
| 101 |
messages.append({"role": "user", "content": build_user_frame(ctx)})
|
| 102 |
|
| 103 |
+
# language steering (kept minimal, avoids confusion)
|
| 104 |
lang = _lang(ctx)
|
| 105 |
messages.append({
|
| 106 |
"role": "user",
|