feat: Implement AI Chat Completions API with support for file search, web search, and Aktenzeichen-based collection lookup
This commit is contained in:
bsiggel
0
steps/ai/__init__.py
Normal file
0
steps/ai/__init__.py
Normal file
529
steps/ai/chat_completions_api_step.py
Normal file
529
steps/ai/chat_completions_api_step.py
Normal file
@@ -0,0 +1,529 @@
|
||||
"""AI Chat Completions API
|
||||
|
||||
Universal OpenAI-compatible Chat Completions API with xAI/LangChain Backend.
|
||||
|
||||
Features:
|
||||
- File Search (RAG) via xAI Collections
|
||||
- Web Search via xAI web_search tool
|
||||
- Aktenzeichen-based automatic collection lookup
|
||||
- Streaming & Non-Streaming support
|
||||
- Multiple tools simultaneously (file_search + web_search)
|
||||
"""
|
||||
import json
|
||||
import time
|
||||
from typing import Any, Dict, List, Optional
|
||||
from motia import FlowContext, http, ApiRequest, ApiResponse
|
||||
|
||||
|
||||
config = {
|
||||
"name": "AI Chat Completions API",
|
||||
"description": "Universal OpenAI-compatible Chat Completions API with xAI backend, RAG, and web search",
|
||||
"flows": ["ai-general"],
|
||||
"triggers": [
|
||||
http("POST", "/ai/chat/completions")
|
||||
],
|
||||
}
|
||||
|
||||
|
||||
async def handler(request: ApiRequest, ctx: FlowContext[Any]) -> ApiResponse:
|
||||
"""
|
||||
OpenAI-compatible Chat Completions endpoint.
|
||||
|
||||
Request Body (OpenAI format):
|
||||
{
|
||||
"model": "grok-4.20-beta-0309-reasoning",
|
||||
"messages": [
|
||||
{"role": "system", "content": "You are helpful"},
|
||||
{"role": "user", "content": "1234/56 Was ist der Stand?"}
|
||||
],
|
||||
"temperature": 0.7,
|
||||
"max_tokens": 2000,
|
||||
"stream": false,
|
||||
"extra_body": {
|
||||
"collection_id": "col_abc123", // Optional: override auto-detection
|
||||
"enable_web_search": true, // Optional: enable web search (default: false)
|
||||
"web_search_config": { // Optional: web search configuration
|
||||
"allowed_domains": ["example.com"],
|
||||
"excluded_domains": ["spam.com"],
|
||||
"enable_image_understanding": true
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
Aktenzeichen-Erkennung (Priority):
|
||||
1. extra_body.collection_id (explicit override)
|
||||
2. First user message starts with Aktenzeichen (e.g., "1234/56 ...")
|
||||
3. Web-only mode if no collection_id (must enable_web_search)
|
||||
|
||||
Response (OpenAI format):
|
||||
Non-Streaming:
|
||||
{
|
||||
"id": "chatcmpl-...",
|
||||
"object": "chat.completion",
|
||||
"created": 1234567890,
|
||||
"model": "grok-4.20-beta-0309-reasoning",
|
||||
"choices": [{
|
||||
"index": 0,
|
||||
"message": {"role": "assistant", "content": "..."},
|
||||
"finish_reason": "stop"
|
||||
}],
|
||||
"usage": {"prompt_tokens": X, "completion_tokens": Y, "total_tokens": Z}
|
||||
}
|
||||
|
||||
Streaming (SSE):
|
||||
data: {"id":"chatcmpl-...","choices":[{"delta":{"content":"Hello"},...}]}
|
||||
data: {"id":"chatcmpl-...","choices":[{"delta":{"content":" world"},...}]}
|
||||
data: {"choices":[{"delta":{},"finish_reason":"stop"}]}
|
||||
data: [DONE]
|
||||
"""
|
||||
from services.langchain_xai_service import LangChainXAIService
|
||||
from services.aktenzeichen_utils import extract_aktenzeichen, normalize_aktenzeichen
|
||||
from services.espocrm import EspoCRMAPI
|
||||
|
||||
ctx.logger.info("=" * 80)
|
||||
ctx.logger.info("🤖 AI CHAT COMPLETIONS API")
|
||||
ctx.logger.info("=" * 80)
|
||||
|
||||
try:
|
||||
# Parse request body
|
||||
body = request.body or {}
|
||||
|
||||
if not isinstance(body, dict):
|
||||
ctx.logger.error(f"❌ Invalid request body type: {type(body)}")
|
||||
return ApiResponse(
|
||||
status=400,
|
||||
body={'error': 'Request body must be JSON object'}
|
||||
)
|
||||
|
||||
# Extract parameters
|
||||
model_name = body.get('model', 'grok-4.20-beta-0309-reasoning')
|
||||
messages = body.get('messages', [])
|
||||
temperature = body.get('temperature', 0.7)
|
||||
max_tokens = body.get('max_tokens')
|
||||
stream = body.get('stream', False)
|
||||
extra_body = body.get('extra_body', {})
|
||||
|
||||
# Web Search parameters (default: disabled)
|
||||
enable_web_search = extra_body.get('enable_web_search', False)
|
||||
web_search_config = extra_body.get('web_search_config', {})
|
||||
|
||||
ctx.logger.info(f"📋 Model: {model_name}")
|
||||
ctx.logger.info(f"📋 Messages: {len(messages)}")
|
||||
ctx.logger.info(f"📋 Stream: {stream}")
|
||||
ctx.logger.info(f"📋 Web Search: {'enabled' if enable_web_search else 'disabled'}")
|
||||
if enable_web_search and web_search_config:
|
||||
ctx.logger.debug(f"Web Search Config: {json.dumps(web_search_config, indent=2)}")
|
||||
|
||||
# Log full conversation messages
|
||||
ctx.logger.info("-" * 80)
|
||||
ctx.logger.info("📨 REQUEST MESSAGES:")
|
||||
for i, msg in enumerate(messages, 1):
|
||||
role = msg.get('role', 'unknown')
|
||||
content = msg.get('content', '')
|
||||
preview = content[:150] + "..." if len(content) > 150 else content
|
||||
ctx.logger.info(f" [{i}] {role}: {preview}")
|
||||
ctx.logger.info("-" * 80)
|
||||
|
||||
# Validate messages
|
||||
if not messages or not isinstance(messages, list):
|
||||
ctx.logger.error("❌ Missing or invalid messages array")
|
||||
return ApiResponse(
|
||||
status=400,
|
||||
body={'error': 'messages must be non-empty array'}
|
||||
)
|
||||
|
||||
# Determine collection_id (Priority: extra_body > Aktenzeichen > optional for web-only)
|
||||
collection_id: Optional[str] = None
|
||||
aktenzeichen: Optional[str] = None
|
||||
|
||||
# Priority 1: Explicit collection_id in extra_body
|
||||
if 'collection_id' in extra_body:
|
||||
collection_id = extra_body['collection_id']
|
||||
ctx.logger.info(f"🔍 Collection ID from extra_body: {collection_id}")
|
||||
|
||||
# Priority 2: Extract Aktenzeichen from first user message
|
||||
else:
|
||||
for msg in messages:
|
||||
if msg.get('role') == 'user':
|
||||
content = msg.get('content', '')
|
||||
aktenzeichen_raw = extract_aktenzeichen(content)
|
||||
|
||||
if aktenzeichen_raw:
|
||||
aktenzeichen = normalize_aktenzeichen(aktenzeichen_raw)
|
||||
ctx.logger.info(f"🔍 Aktenzeichen detected: {aktenzeichen}")
|
||||
|
||||
# Lookup collection_id via EspoCRM
|
||||
collection_id = await lookup_collection_by_aktenzeichen(
|
||||
aktenzeichen, ctx
|
||||
)
|
||||
|
||||
if collection_id:
|
||||
ctx.logger.info(f"✅ Collection found: {collection_id}")
|
||||
|
||||
# Remove Aktenzeichen from message (clean prompt)
|
||||
from services.aktenzeichen_utils import remove_aktenzeichen
|
||||
msg['content'] = remove_aktenzeichen(content)
|
||||
ctx.logger.debug(f"Cleaned message: {msg['content']}")
|
||||
else:
|
||||
ctx.logger.warn(f"⚠️ No collection found for {aktenzeichen}")
|
||||
|
||||
break # Only check first user message
|
||||
|
||||
# Priority 3: Error if no collection_id AND web_search disabled
|
||||
if not collection_id and not enable_web_search:
|
||||
ctx.logger.error("❌ No collection_id found and web_search disabled")
|
||||
ctx.logger.error(" Provide collection_id, enable web_search, or both")
|
||||
return ApiResponse(
|
||||
status=400,
|
||||
body={
|
||||
'error': 'collection_id or web_search required',
|
||||
'message': 'Provide collection_id in extra_body, enable web_search, or start message with Aktenzeichen (e.g., "1234/56 question")'
|
||||
}
|
||||
)
|
||||
|
||||
# Initialize LangChain xAI Service
|
||||
try:
|
||||
langchain_service = LangChainXAIService(ctx)
|
||||
except ValueError as e:
|
||||
ctx.logger.error(f"❌ Service initialization failed: {e}")
|
||||
return ApiResponse(
|
||||
status=500,
|
||||
body={'error': 'Service configuration error', 'details': str(e)}
|
||||
)
|
||||
|
||||
# Create ChatXAI model
|
||||
model = langchain_service.get_chat_model(
|
||||
model=model_name,
|
||||
temperature=temperature,
|
||||
max_tokens=max_tokens
|
||||
)
|
||||
|
||||
# Bind tools (file_search and/or web_search)
|
||||
model_with_tools = langchain_service.bind_tools(
|
||||
model=model,
|
||||
collection_id=collection_id,
|
||||
enable_web_search=enable_web_search,
|
||||
web_search_config=web_search_config,
|
||||
max_num_results=10
|
||||
)
|
||||
|
||||
# Generate completion_id
|
||||
completion_id = f"chatcmpl-{ctx.traceId[:12]}" if hasattr(ctx, 'traceId') else f"chatcmpl-{int(time.time())}"
|
||||
created_ts = int(time.time())
|
||||
|
||||
# Branch: Streaming vs Non-Streaming
|
||||
if stream:
|
||||
ctx.logger.info("🌊 Starting streaming response...")
|
||||
return await handle_streaming_response(
|
||||
model_with_tools=model_with_tools,
|
||||
messages=messages,
|
||||
completion_id=completion_id,
|
||||
created_ts=created_ts,
|
||||
model_name=model_name,
|
||||
langchain_service=langchain_service,
|
||||
ctx=ctx
|
||||
)
|
||||
else:
|
||||
ctx.logger.info("📦 Starting non-streaming response...")
|
||||
return await handle_non_streaming_response(
|
||||
model_with_tools=model_with_tools,
|
||||
messages=messages,
|
||||
completion_id=completion_id,
|
||||
created_ts=created_ts,
|
||||
model_name=model_name,
|
||||
langchain_service=langchain_service,
|
||||
ctx=ctx
|
||||
)
|
||||
|
||||
except Exception as e:
|
||||
ctx.logger.error("=" * 80)
|
||||
ctx.logger.error("❌ ERROR: AI CHAT COMPLETIONS API")
|
||||
ctx.logger.error("=" * 80)
|
||||
ctx.logger.error(f"Error: {e}", exc_info=True)
|
||||
ctx.logger.error(f"Request body: {json.dumps(request.body, indent=2, ensure_ascii=False)}")
|
||||
ctx.logger.error("=" * 80)
|
||||
|
||||
return ApiResponse(
|
||||
status=500,
|
||||
body={
|
||||
'error': 'Internal server error',
|
||||
'message': str(e)
|
||||
}
|
||||
)
|
||||
|
||||
|
||||
async def handle_non_streaming_response(
|
||||
model_with_tools,
|
||||
messages: List[Dict[str, Any]],
|
||||
completion_id: str,
|
||||
created_ts: int,
|
||||
model_name: str,
|
||||
langchain_service,
|
||||
ctx: FlowContext
|
||||
) -> ApiResponse:
|
||||
"""
|
||||
Handle non-streaming chat completion.
|
||||
|
||||
Returns:
|
||||
ApiResponse with OpenAI-format JSON body
|
||||
"""
|
||||
try:
|
||||
# Invoke model
|
||||
result = await langchain_service.invoke_chat(model_with_tools, messages)
|
||||
|
||||
# Extract content - handle both string and structured responses
|
||||
if hasattr(result, 'content'):
|
||||
raw_content = result.content
|
||||
|
||||
# If content is a list (tool calls + text message), extract text
|
||||
if isinstance(raw_content, list):
|
||||
# Find the text message (usually last element with type='text')
|
||||
text_messages = [
|
||||
item.get('text', '')
|
||||
for item in raw_content
|
||||
if isinstance(item, dict) and item.get('type') == 'text'
|
||||
]
|
||||
content = text_messages[0] if text_messages else str(raw_content)
|
||||
else:
|
||||
content = raw_content
|
||||
else:
|
||||
content = str(result)
|
||||
|
||||
# Build OpenAI-compatible response
|
||||
response_body = {
|
||||
'id': completion_id,
|
||||
'object': 'chat.completion',
|
||||
'created': created_ts,
|
||||
'model': model_name,
|
||||
'choices': [{
|
||||
'index': 0,
|
||||
'message': {
|
||||
'role': 'assistant',
|
||||
'content': content
|
||||
},
|
||||
'finish_reason': 'stop'
|
||||
}],
|
||||
'usage': {
|
||||
'prompt_tokens': 0, # LangChain doesn't expose token counts easily
|
||||
'completion_tokens': 0,
|
||||
'total_tokens': 0
|
||||
}
|
||||
}
|
||||
|
||||
# Log token usage (if available)
|
||||
if hasattr(result, 'usage_metadata'):
|
||||
usage = result.usage_metadata
|
||||
prompt_tokens = getattr(usage, 'input_tokens', 0)
|
||||
completion_tokens = getattr(usage, 'output_tokens', 0)
|
||||
response_body['usage'] = {
|
||||
'prompt_tokens': prompt_tokens,
|
||||
'completion_tokens': completion_tokens,
|
||||
'total_tokens': prompt_tokens + completion_tokens
|
||||
}
|
||||
ctx.logger.info(f"📊 Token Usage: prompt={prompt_tokens}, completion={completion_tokens}")
|
||||
|
||||
# Log citations if available (from tool response annotations)
|
||||
if hasattr(result, 'content') and isinstance(result.content, list):
|
||||
# Extract citations from structured response
|
||||
for item in result.content:
|
||||
if isinstance(item, dict) and item.get('type') == 'text':
|
||||
annotations = item.get('annotations', [])
|
||||
if annotations:
|
||||
ctx.logger.info(f"🔗 Citations: {len(annotations)}")
|
||||
for i, citation in enumerate(annotations[:10], 1): # Log first 10
|
||||
url = citation.get('url', 'N/A')
|
||||
title = citation.get('title', '')
|
||||
if url.startswith('collections://'):
|
||||
# Internal collection reference
|
||||
ctx.logger.debug(f" [{i}] Collection Document: {title}")
|
||||
else:
|
||||
# External URL
|
||||
ctx.logger.debug(f" [{i}] {url}")
|
||||
|
||||
# Log complete response content
|
||||
ctx.logger.info(f"✅ Chat completion: {len(content)} chars")
|
||||
ctx.logger.info("=" * 80)
|
||||
ctx.logger.info("📝 COMPLETE RESPONSE:")
|
||||
ctx.logger.info("-" * 80)
|
||||
ctx.logger.info(content)
|
||||
ctx.logger.info("-" * 80)
|
||||
ctx.logger.info("=" * 80)
|
||||
|
||||
return ApiResponse(
|
||||
status=200,
|
||||
body=response_body
|
||||
)
|
||||
|
||||
except Exception as e:
|
||||
ctx.logger.error(f"❌ Non-streaming completion failed: {e}", exc_info=True)
|
||||
raise
|
||||
|
||||
|
||||
async def handle_streaming_response(
|
||||
model_with_tools,
|
||||
messages: List[Dict[str, Any]],
|
||||
completion_id: str,
|
||||
created_ts: int,
|
||||
model_name: str,
|
||||
langchain_service,
|
||||
ctx: FlowContext
|
||||
):
|
||||
"""
|
||||
Handle streaming chat completion via SSE.
|
||||
|
||||
Returns:
|
||||
Streaming response generator
|
||||
"""
|
||||
async def stream_generator():
|
||||
try:
|
||||
# Set SSE headers
|
||||
await ctx.response.status(200)
|
||||
await ctx.response.headers({
|
||||
"Content-Type": "text/event-stream",
|
||||
"Cache-Control": "no-cache",
|
||||
"Connection": "keep-alive"
|
||||
})
|
||||
|
||||
ctx.logger.info("🌊 Streaming started")
|
||||
|
||||
# Stream chunks
|
||||
chunk_count = 0
|
||||
total_content = ""
|
||||
|
||||
async for chunk in langchain_service.astream_chat(model_with_tools, messages):
|
||||
# Extract delta content - handle structured chunks
|
||||
if hasattr(chunk, "content"):
|
||||
chunk_content = chunk.content
|
||||
|
||||
# If chunk content is a list (tool calls), extract text parts
|
||||
if isinstance(chunk_content, list):
|
||||
# Accumulate only text deltas
|
||||
text_parts = [
|
||||
item.get('text', '')
|
||||
for item in chunk_content
|
||||
if isinstance(item, dict) and item.get('type') == 'text'
|
||||
]
|
||||
delta = ''.join(text_parts)
|
||||
else:
|
||||
delta = chunk_content
|
||||
else:
|
||||
delta = ""
|
||||
|
||||
if delta:
|
||||
total_content += delta
|
||||
chunk_count += 1
|
||||
|
||||
# Build SSE data
|
||||
data = {
|
||||
"id": completion_id,
|
||||
"object": "chat.completion.chunk",
|
||||
"created": created_ts,
|
||||
"model": model_name,
|
||||
"choices": [{
|
||||
"index": 0,
|
||||
"delta": {"content": delta},
|
||||
"finish_reason": None
|
||||
}]
|
||||
}
|
||||
|
||||
# Send SSE event
|
||||
await ctx.response.stream(f"data: {json.dumps(data, ensure_ascii=False)}\n\n")
|
||||
|
||||
# Send finish event
|
||||
finish_data = {
|
||||
"id": completion_id,
|
||||
"object": "chat.completion.chunk",
|
||||
"created": created_ts,
|
||||
"model": model_name,
|
||||
"choices": [{
|
||||
"index": 0,
|
||||
"delta": {},
|
||||
"finish_reason": "stop"
|
||||
}]
|
||||
}
|
||||
await ctx.response.stream(f"data: {json.dumps(finish_data)}\n\n")
|
||||
|
||||
# Send [DONE]
|
||||
await ctx.response.stream("data: [DONE]\n\n")
|
||||
|
||||
# Close stream
|
||||
await ctx.response.close()
|
||||
|
||||
# Log complete streamed response
|
||||
ctx.logger.info(f"✅ Streaming completed: {chunk_count} chunks, {len(total_content)} chars")
|
||||
ctx.logger.info("=" * 80)
|
||||
ctx.logger.info("📝 COMPLETE STREAMED RESPONSE:")
|
||||
ctx.logger.info("-" * 80)
|
||||
ctx.logger.info(total_content)
|
||||
ctx.logger.info("-" * 80)
|
||||
ctx.logger.info("=" * 80)
|
||||
|
||||
except Exception as e:
|
||||
ctx.logger.error(f"❌ Streaming failed: {e}", exc_info=True)
|
||||
|
||||
# Send error event
|
||||
error_data = {
|
||||
"error": {
|
||||
"message": str(e),
|
||||
"type": "server_error"
|
||||
}
|
||||
}
|
||||
await ctx.response.stream(f"data: {json.dumps(error_data)}\n\n")
|
||||
await ctx.response.close()
|
||||
|
||||
return stream_generator()
|
||||
|
||||
|
||||
async def lookup_collection_by_aktenzeichen(
|
||||
aktenzeichen: str,
|
||||
ctx: FlowContext
|
||||
) -> Optional[str]:
|
||||
"""
|
||||
Lookup xAI Collection ID for Aktenzeichen via EspoCRM.
|
||||
|
||||
Search strategy:
|
||||
1. Search for Raeumungsklage with matching advowareAkteBezeichner
|
||||
2. Return xaiCollectionId if found
|
||||
|
||||
Args:
|
||||
aktenzeichen: Normalized Aktenzeichen (e.g., "1234/56")
|
||||
ctx: Motia context
|
||||
|
||||
Returns:
|
||||
Collection ID or None if not found
|
||||
"""
|
||||
try:
|
||||
# Initialize EspoCRM API
|
||||
espocrm = EspoCRMAPI(ctx)
|
||||
|
||||
# Search Räumungsklage by advowareAkteBezeichner
|
||||
ctx.logger.info(f"🔍 Searching Räumungsklage for Aktenzeichen: {aktenzeichen}")
|
||||
|
||||
search_result = await espocrm.search_entities(
|
||||
entity_type='Raeumungsklage',
|
||||
where=[{
|
||||
'type': 'equals',
|
||||
'attribute': 'advowareAkteBezeichner',
|
||||
'value': aktenzeichen
|
||||
}],
|
||||
select=['id', 'xaiCollectionId', 'advowareAkteBezeichner'],
|
||||
maxSize=1
|
||||
)
|
||||
|
||||
if search_result and len(search_result) > 0:
|
||||
entity = search_result[0]
|
||||
collection_id = entity.get('xaiCollectionId')
|
||||
|
||||
if collection_id:
|
||||
ctx.logger.info(f"✅ Found Räumungsklage: {entity.get('id')}")
|
||||
return collection_id
|
||||
else:
|
||||
ctx.logger.warn(f"⚠️ Räumungsklage found but no xaiCollectionId: {entity.get('id')}")
|
||||
else:
|
||||
ctx.logger.warn(f"⚠️ No Räumungsklage found for {aktenzeichen}")
|
||||
|
||||
return None
|
||||
|
||||
except Exception as e:
|
||||
ctx.logger.error(f"❌ Collection lookup failed: {e}", exc_info=True)
|
||||
return None
|
||||
Reference in New Issue
Block a user