Sophomores's picture
fix(multimodal): remove fake image support, model cannot decode base64
ad06fde
Raw
History Blame Contribute Delete
11.6 kB
"""Tool calling and multimodal message parsing."""
import json
import re
import uuid
import base64
import io
MAX_IMAGE_B64_SIZE = 50000 # ~37KB raw image
def _compress_b64_if_needed(b64: str) -> str:
"""Compress image if base64 is too large for text embedding."""
if len(b64) <= MAX_IMAGE_B64_SIZE:
return b64
try:
from PIL import Image
img_data = base64.b64decode(b64)
img = Image.open(io.BytesIO(img_data))
# Resize to max 256px on longest side
max_dim = 256
ratio = min(max_dim / img.width, max_dim / img.height)
if ratio < 1:
img = img.resize((int(img.width * ratio), int(img.height * ratio)), Image.LANCZOS)
# Convert to JPEG with quality reduction
buf = io.BytesIO()
img.convert("RGB").save(buf, format="JPEG", quality=60)
compressed = base64.b64encode(buf.getvalue()).decode()
return compressed
except Exception:
# If PIL not available, truncate (model will get partial data)
return b64[:MAX_IMAGE_B64_SIZE]
def _build_tool_choice_instruction(tool_choice, tool_defs: list) -> str:
"""Build tool_choice constraint instruction.
tool_choice values:
- "none": do not call any tool
- "auto": decide whether to call tools (default)
- "required": must call at least one tool
- {"type": "function", "function": {"name": "xxx"}}: must call specific tool
"""
if tool_choice == "none":
return "\n\nIMPORTANT: Do NOT call any tools. Respond with text only."
if tool_choice == "required":
return "\n\nIMPORTANT: You MUST call at least one tool. Do not respond with text only."
if isinstance(tool_choice, dict):
fn_name = tool_choice.get("function", {}).get("name", "")
if fn_name:
return f'\n\nIMPORTANT: You MUST call the tool "{fn_name}". Do not call other tools.'
return ""
def messages_to_prompt(messages: list, tools: list = None, tool_choice=None) -> tuple:
"""Convert OpenAI messages to (prompt_str, images_list).
Returns (prompt, images) where images is a list of (bytes, mime_type) tuples.
"""
parts = []
images = []
if tools and tool_choice != "none":
tool_defs = []
for tool in tools:
fn = tool.get("function", tool) if tool.get("type") == "function" else tool
tool_defs.append({
"name": fn.get("name", tool.get("name", "")),
"description": fn.get("description", tool.get("description", "")),
"parameters": fn.get("parameters", tool.get("parameters", {})),
})
if tool_defs:
constraint = _build_tool_choice_instruction(tool_choice, tool_defs)
parts.append(
"# Tool Use\n\n"
"You can call the following tools. Call format:\n"
'```tool_call\n{"name": "func_name", "arguments": {...}}\n```\n'
"When calling tools, output ONLY the tool_call block(s).\n\n"
f"Available tools:\n{json.dumps(tool_defs, indent=2)}"
f"{constraint}"
)
for msg in messages:
role = msg.get("role", "user")
content = msg.get("content", "")
if isinstance(content, list):
text_parts = []
for c in content:
if c.get("type") in ("text", "input_text"):
text_parts.append(c.get("text", ""))
elif c.get("type") == "image_url":
text_parts.append("[Note: Image input not supported in this API. Please describe the image in text.]")
elif c.get("type") == "image":
text_parts.append("[Note: Image input not supported in this API. Please describe the image in text.]")
content = " ".join(text_parts)
if role == "system":
parts.append(f"[System instruction]: {content}")
elif role == "assistant":
if msg.get("tool_calls"):
tc_strs = []
for tc in msg["tool_calls"]:
fn = tc.get("function", {})
tc_strs.append(
f'```tool_call\n{{"name": "{fn.get("name")}", '
f'"arguments": {fn.get("arguments", "{}")}}}\n```'
)
parts.append(f"[Assistant]: {content or ''}\n" + "\n".join(tc_strs))
else:
parts.append(f"[Assistant]: {content}")
elif role == "tool":
parts.append(f"[Tool result for {msg.get('name', '')}]: {content}")
else:
parts.append(content if content else "")
prompt = "\n\n".join(p for p in parts if p)
return prompt, images
def parse_tool_calls(text: str) -> tuple:
"""Extract tool_call blocks. Returns (clean_text, tool_calls_list)."""
tool_calls = []
pattern = r'```tool_call\s*\n(.*?)\n```'
clean_parts = []
last_end = 0
for m in re.finditer(pattern, text, re.DOTALL):
clean_parts.append(text[last_end:m.start()])
last_end = m.end()
try:
data = json.loads(m.group(1).strip())
tool_calls.append({
"id": f"call_{uuid.uuid4().hex[:8]}",
"type": "function",
"function": {
"name": data["name"],
"arguments": json.dumps(data.get("arguments", {}), ensure_ascii=False),
},
})
except (json.JSONDecodeError, KeyError):
pass
clean_parts.append(text[last_end:])
clean = "".join(clean_parts).strip()
return clean, tool_calls
# ─── Google Native API helpers ─────────────────────────────────────────────────
def build_tool_prompt(tool_defs: list) -> str:
"""Build natural tool-use prompt for Gemini Web that avoids prompt-injection detection."""
tool_spec = json.dumps(tool_defs, indent=2, ensure_ascii=False)
return (
"# Tool Use\n\n"
"You can call the following tools to help accomplish tasks. "
"These tools connect to the user's local environment and will execute when called.\n\n"
"Call format (use this exact format):\n"
"```function_call\n"
'{"name": "<tool_name>", "args": {<arguments>}}\n'
"```\n\n"
"When calling tools:\n"
"- Output ONLY the function_call block(s), nothing else\n"
"- You may call multiple tools with multiple blocks\n"
"- After receiving a [Tool result for ...], use that data to answer the user\n\n"
f"Available tools:\n{tool_spec}"
)
def _google_tool_choice_instruction(req: dict) -> str:
"""Extract tool_choice constraint from Google API toolConfig."""
tool_config = req.get("toolConfig", {})
fc_config = tool_config.get("functionCallingConfig", {})
mode = fc_config.get("mode", "AUTO")
allowed = fc_config.get("allowedFunctionNames", [])
if mode == "NONE":
return "\n\nIMPORTANT: Do NOT call any tools. Respond with text only."
if mode == "ANY":
if allowed:
names = ", ".join(f'"{n}"' for n in allowed)
return f"\n\nIMPORTANT: You MUST call one of these tools: {names}. Do not respond with text only."
return "\n\nIMPORTANT: You MUST call at least one tool. Do not respond with text only."
return ""
def google_contents_to_prompt(req: dict) -> tuple:
"""Convert Google API contents/tools/systemInstruction to (prompt_str, images_list).
Returns (prompt, images) where images is a list of (bytes, mime_type) tuples.
"""
parts = []
images = []
tool_config = req.get("toolConfig", {})
fc_mode = tool_config.get("functionCallingConfig", {}).get("mode", "AUTO")
tools = req.get("tools")
tool_defs = []
if tools and fc_mode != "NONE":
for tool_group in tools:
for fn in tool_group.get("functionDeclarations", []):
td = {"name": fn.get("name", ""), "description": fn.get("description", "")}
params = fn.get("parameters") or fn.get("parametersJsonSchema")
if params:
td["parameters"] = params
tool_defs.append(td)
sys_inst = req.get("systemInstruction")
if sys_inst:
sys_parts = sys_inst.get("parts", [])
sys_text = " ".join(p.get("text", "") for p in sys_parts if p.get("text"))
if sys_text:
if tool_defs:
constraint = _google_tool_choice_instruction(req)
parts.append(sys_text + "\n\n" + build_tool_prompt(tool_defs) + constraint)
else:
parts.append(sys_text)
elif tool_defs:
constraint = _google_tool_choice_instruction(req)
parts.append(build_tool_prompt(tool_defs) + constraint)
for content in req.get("contents", []):
role = content.get("role", "user")
msg_parts = []
for p in content.get("parts", []):
if p.get("text"):
msg_parts.append(p["text"])
elif p.get("inlineData"):
data = p["inlineData"]
mime = data.get("mimeType", "image/png")
images.append((base64.b64decode(data["data"]), mime))
elif p.get("functionCall"):
fc = p["functionCall"]
msg_parts.append(
f'```function_call\n{json.dumps({"name": fc["name"], "args": fc.get("args", {})}, ensure_ascii=False)}\n```'
)
elif p.get("functionResponse"):
fr = p["functionResponse"]
msg_parts.append(
f'[Tool result for {fr.get("name", "")}]: {json.dumps(fr.get("response", {}), ensure_ascii=False)}'
)
text = "\n".join(msg_parts)
if role == "model":
parts.append(f"[Assistant]: {text}")
else:
parts.append(text)
return "\n\n".join(p for p in parts if p), images
def parse_google_function_calls(text: str) -> tuple:
"""Extract function_call blocks from model output.
Handles 3 formats:
1. ```function_call\\n{...}\\n``` (standard)
2. function_call\\n{...} (without backticks)
3. Raw JSON with "name" + "args" keys
Returns (clean_text, [{"name": ..., "args": ...}])
"""
function_calls = []
pattern1 = r'```function_call\s*\n(.*?)\n```'
pattern2 = r'(?:^|\n)function_call\s*\n(\{[^`]*?\})'
clean = text
for pattern in [pattern1, pattern2]:
for match in re.findall(pattern, clean, re.DOTALL):
try:
data = json.loads(match.strip())
if "name" in data:
function_calls.append({
"name": data["name"],
"args": data.get("args", data.get("arguments", {})),
})
except (json.JSONDecodeError, KeyError):
pass
clean = re.sub(pattern, '', clean, flags=re.DOTALL).strip()
if not function_calls and clean.strip().startswith("{"):
try:
data = json.loads(clean.strip())
if "name" in data and ("args" in data or "arguments" in data):
function_calls.append({
"name": data["name"],
"args": data.get("args", data.get("arguments", {})),
})
clean = ""
except (json.JSONDecodeError, KeyError):
pass
return clean, function_calls