sunholo 0.63.0__tar.gz → 0.64.4__tar.gz

{sunholo-0.63.0 → sunholo-0.64.4}/PKG-INFO

@@ -1,9 +1,9 @@
 Metadata-Version: 2.1
 Name: sunholo
-Version: 0.63.0
+Version: 0.64.4
 Summary: Large Language Model DevOps - a package to help deploy LLMs to the Cloud.
 Home-page: https://github.com/sunholo-data/sunholo-py
-Download-URL: https://github.com/sunholo-data/sunholo-py/archive/refs/tags/v0.63.0.tar.gz
+Download-URL: https://github.com/sunholo-data/sunholo-py/archive/refs/tags/v0.64.4.tar.gz
 Author: Holosun ApS
 Author-email: multivac@sunholo.com
 License: Apache License, Version 2.0
@@ -19,6 +19,7 @@ Classifier: Programming Language :: Python :: 3.12
 Description-Content-Type: text/markdown
 License-File: LICENSE.txt
 Requires-Dist: google-auth
+Requires-Dist: tenacity==8.3.0
 Requires-Dist: langchain
 Requires-Dist: langchain_experimental
 Requires-Dist: langchain-community

{sunholo-0.63.0 → sunholo-0.64.4}/setup.py

@@ -1,7 +1,7 @@
 from setuptools import setup, find_packages
 
 # Define your base version
-version = '0.63.0'
+version = '0.64.4'
 
 setup(
     name='sunholo',
@@ -29,6 +29,7 @@ setup(
     install_requires=[
         # Base dependencies
         "google-auth", # to check if on gcp
+        "tenacity==8.3.0",
         "langchain",
         "langchain_experimental",
         "langchain-community",

{sunholo-0.63.0 → sunholo-0.64.4}/sunholo/agents/flask/qna_routes.py

@@ -16,6 +16,7 @@
 import json
 import traceback
 import datetime
+import uuid
 
 from ...agents import extract_chat_history, handle_special_commands
 from ...qna.parsers import parse_output
@@ -23,7 +24,11 @@ from ...streaming import start_streaming_chat
 from ...archive import archive_qa
 from ...logging import log
 from ...utils.config import load_config
- 
+from ...utils.version import sunholo_version
+import os
+from ...gcs.add_file import add_file_to_gcs
+
+
 try:
     from flask import request, jsonify, Response
 except ImportError:
@@ -34,6 +39,7 @@ try:
 except ImportError:
     pass    
 
+
 def register_qna_routes(app, stream_interpreter, vac_interpreter):
 
     @app.route("/")
@@ -155,9 +161,122 @@ def register_qna_routes(app, stream_interpreter, vac_interpreter):
             span.end(output=jsonify(bot_output))
             trace.update(output=jsonify(bot_output)) 
 
+        # {'answer': 'output'}
         return jsonify(bot_output)
 
-    # Any other QNA related routes can be added here
+    @app.route('/openai/v1/chat/completions', methods=['POST'])
+    @app.route('/openai/v1/chat/completions/<vector_name>', methods=['POST'])
+    def openai_compatible_endpoint(vector_name=None):
+        data = request.get_json()
+        log.info(f'openai_compatible_endpoint got data: {data}')
+
+        vector_name = vector_name or data.pop('model', None)
+        messages = data.pop('messages', None)
+        chat_history = data.pop('chat_history', None)
+        stream = data.pop('stream', False)
+
+        log.info(f'openai_compatible_endpoint got data: {data} for vector: {vector_name}')
+
+        if not messages:
+            return jsonify({"error": "No messages provided"}), 400
+
+        user_message = next((msg['content'] for msg in messages if msg['role'] == 'user'), None)
+        if not user_message:
+            return jsonify({"error": "No user message provided"}), 400
+
+        all_input = {
+            "user_input": user_message,
+            "chat_history": chat_history,
+            "kwargs": data
+        }
+
+        observed_stream_interpreter = observe()(stream_interpreter)
+
+
+        response_id = str(uuid.uuid4())
+        
+        def generate_response_content():
+            for chunk in start_streaming_chat(question=user_message,
+                                            vector_name=vector_name,
+                                            qna_func=observed_stream_interpreter,
+                                            chat_history=all_input["chat_history"],
+                                            wait_time=all_input.get("stream_wait_time", 1),
+                                            timeout=all_input.get("stream_timeout", 60),
+                                            **all_input["kwargs"]
+                                            ):
+                if isinstance(chunk, dict) and 'answer' in chunk:
+                    openai_chunk = {
+                        "id": response_id,
+                        "object": "chat.completion.chunk",
+                        "created": str(datetime.time()),
+                        "model": vector_name,
+                        "system_fingerprint": sunholo_version(),
+                        "choices": [{
+                            "index": 0,
+                            "delta": {"content": chunk['answer']},
+                            "logprobs": None,
+                            "finish_reason": None
+                        }]
+                    }
+                    yield json.dumps(openai_chunk) + "\n"
+                else:
+                    log.info(f"Unknown chunk: {chunk}")
+
+            final_chunk = {
+                "id": response_id,
+                "object": "chat.completion.chunk",
+                "created": str(datetime.time()),
+                "model": vector_name,
+                "system_fingerprint": sunholo_version(),
+                "choices": [{
+                    "index": 0,
+                    "delta": {},
+                    "logprobs": None,
+                    "finish_reason": "stop"
+                }]
+            }
+            yield json.dumps(final_chunk) + "\n"
+
+        if stream:
+            return Response(generate_response_content(), content_type='text/plain; charset=utf-8')
+
+        try:
+            bot_output = observed_stream_interpreter(
+                question=user_message,
+                vector_name=vector_name,
+                chat_history=all_input["chat_history"],
+                **all_input["kwargs"]
+            )
+            bot_output = parse_output(bot_output)
+
+            openai_response = {
+                "id": response_id,
+                "object": "chat.completion",
+                "created": str(datetime.time()),
+                "model": vector_name,
+                "system_fingerprint": sunholo_version(),
+                "choices": [{
+                    "index": 0,
+                    "message": {
+                        "role": "assistant",
+                        "content": bot_output.get('answer', ''),
+                    },
+                    "logprobs": None,
+                    "finish_reason": "stop"
+                }],
+                "usage": {
+                    "prompt_tokens": len(user_message.split()),
+                    "completion_tokens": len(bot_output.get('answer', '').split()),
+                    "total_tokens": len(user_message.split()) + len(bot_output.get('answer', '').split())
+                }
+            }
+
+            log.info(f"OpenAI response: {openai_response}")
+            return jsonify(openai_response)
+
+        except Exception as err:
+            return jsonify({'error': f'QNA_ERROR: An error occurred: {str(err)} traceback: {traceback.format_exc()}'}), 500
+
 
 def create_langfuse_trace(request, vector_name):
     try:
@@ -172,10 +291,8 @@ def create_langfuse_trace(request, vector_name):
     session_id = request.headers.get("X-Session-ID")
     message_source = request.headers.get("X-Message-Source")
 
-    # can't import tags yet via CallbackHandler
-    from importlib.metadata import version
-    package_version = version('sunholo')
-    tags = [f"sunholo-v{package_version}"]
+    package_version = sunholo_version()
+    tags = [package_version]
     if message_source:
         tags.append(message_source)
 
@@ -191,8 +308,28 @@ def prep_vac(request, vector_name):
     #trace = create_langfuse_trace(request, vector_name)
     trace = None
     span = None
-    data = request.get_json()
+
+    if request.content_type.startswith('application/json'):
+        data = request.get_json()
+    elif request.content_type.startswith('multipart/form-data'):
+        data = request.form.to_dict()
+        if 'file' in request.files:
+            file = request.files['file']
+            if file.filename != '':
+                log.info(f"Found file: {file.filename} to upload to GCS")
+                try:
+                    image_uri, mime_type = handle_file_upload(file, vector_name)
+                    data["image_uri"] = image_uri
+                    data["mime"] = mime_type
+                except Exception as e:
+                    return jsonify({'error': str(e), 'traceback': traceback.format_exc()}), 500
+            else:
+                return jsonify({"error": "No file selected"}), 400
+    else:
+        return jsonify({"error": "Unsupported content type"}), 400
+
     log.info(f"vac/{vector_name} got data: {data}")
+
     config, _ = load_config("config/llm_config.yaml")
     vac_configs = config.get("vac")
     if vac_configs:
@@ -213,7 +350,7 @@ def prep_vac(request, vector_name):
                  'vector_name': vector_name, 
                  'chat_history': paired_messages, 
                  'stream_wait_time': stream_wait_time,
-                 'stream_timeout':stream_timeout,
+                 'stream_timeout': stream_timeout,
                  'kwargs': data}
 
     if trace:
@@ -233,4 +370,14 @@ def prep_vac(request, vector_name):
         "command_response": command_response,
         "all_input": all_input,
         "vac_config": vac_config
-    }
+    }
+
+
+def handle_file_upload(file, vector_name):
+    try:
+        file.save(file.filename)
+        image_uri = add_file_to_gcs(file.filename, vector_name)
+        os.remove(file.filename)  # Clean up the saved file
+        return image_uri, file.mimetype
+    except Exception as e:
+        raise Exception(f'File upload failed: {str(e)}')

{sunholo-0.63.0 → sunholo-0.64.4}/sunholo/langfuse/callback.py

@@ -7,6 +7,8 @@ try:
 except ImportError:
     CallbackHandler = None
 
+from ..utils.version import sunholo_version
+
 def create_langfuse_callback(**kwargs):
 
     if not CallbackHandler:
@@ -40,10 +42,7 @@ def add_langfuse_tracing(
     session_id = request.headers.get("X-Session-ID")
     message_source = request.headers.get("X-Message-Source")
 
-    # can't import tags yet via CallbackHandler
-    from importlib.metadata import version
-    package_version = version('sunholo')
-    tags = [f"sunholo-v{package_version}"]
+    tags = [sunholo_version()]
     if message_source:
         tags.append(message_source)
 

sunholo-0.64.4/sunholo/utils/version.py

@@ -0,0 +1,3 @@
+def sunholo_version():
+    from importlib.metadata import version
+    return f"sunholo-{version('sunholo')}"

{sunholo-0.63.0 → sunholo-0.64.4}/sunholo.egg-info/PKG-INFO

@@ -1,9 +1,9 @@
 Metadata-Version: 2.1
 Name: sunholo
-Version: 0.63.0
+Version: 0.64.4
 Summary: Large Language Model DevOps - a package to help deploy LLMs to the Cloud.
 Home-page: https://github.com/sunholo-data/sunholo-py
-Download-URL: https://github.com/sunholo-data/sunholo-py/archive/refs/tags/v0.63.0.tar.gz
+Download-URL: https://github.com/sunholo-data/sunholo-py/archive/refs/tags/v0.64.4.tar.gz
 Author: Holosun ApS
 Author-email: multivac@sunholo.com
 License: Apache License, Version 2.0
@@ -19,6 +19,7 @@ Classifier: Programming Language :: Python :: 3.12
 Description-Content-Type: text/markdown
 License-File: LICENSE.txt
 Requires-Dist: google-auth
+Requires-Dist: tenacity==8.3.0
 Requires-Dist: langchain
 Requires-Dist: langchain_experimental
 Requires-Dist: langchain-community

{sunholo-0.63.0 → sunholo-0.64.4}/sunholo.egg-info/SOURCES.txt

@@ -108,6 +108,7 @@ sunholo/utils/gcp_project.py
 sunholo/utils/parsers.py
 sunholo/utils/timedelta.py
 sunholo/utils/user_ids.py
+sunholo/utils/version.py
 sunholo/vertex/__init__.py
 sunholo/vertex/init.py
 sunholo/vertex/memory_tools.py

{sunholo-0.63.0 → sunholo-0.64.4}/sunholo.egg-info/requires.txt

@@ -1,4 +1,5 @@
 google-auth
+tenacity==8.3.0
 langchain
 langchain_experimental
 langchain-community