khoj 1.42.3.dev1__py3-none-any.whl → 1.42.4__py3-none-any.whl

khoj/processor/conversation/openai/utils.py

@@ -14,6 +14,7 @@ from openai.lib.streaming.chat import (
     ChatCompletionStreamEvent,
     ContentDeltaEvent,
 )
+from openai.types.chat.chat_completion import ChatCompletion
 from openai.types.chat.chat_completion_chunk import (
     ChatCompletionChunk,
     Choice,
@@ -78,7 +79,11 @@ def completion_with_backoff(
         client = get_openai_client(openai_api_key, api_base_url)
         openai_clients[client_key] = client
 
+    stream = not is_non_streaming_model(model_name, api_base_url)
     stream_processor = default_stream_processor
+    if stream:
+        model_kwargs["stream_options"] = {"include_usage": True}
+
     formatted_messages = format_message_for_api(messages, api_base_url)
 
     # Tune reasoning models arguments
@@ -105,27 +110,37 @@ def completion_with_backoff(
         stream_processor = partial(in_stream_thought_processor, thought_tag="think")
         # Reasoning is enabled by default. Disable when deepthought is False.
         # See https://qwenlm.github.io/blog/qwen3/#advanced-usages
-        if not deepthought and len(formatted_messages) > 0:
-            formatted_messages[-1]["content"] = formatted_messages[-1]["content"] + " /no_think"
+        if not deepthought:
+            add_qwen_no_think_tag(formatted_messages)
 
     read_timeout = 300 if is_local_api(api_base_url) else 60
-    model_kwargs["stream_options"] = {"include_usage": True}
     if os.getenv("KHOJ_LLM_SEED"):
         model_kwargs["seed"] = int(os.getenv("KHOJ_LLM_SEED"))
 
     aggregated_response = ""
-    with client.beta.chat.completions.stream(
-        messages=formatted_messages,  # type: ignore
-        model=model_name,
-        temperature=temperature,
-        timeout=httpx.Timeout(30, read=read_timeout),
-        **model_kwargs,
-    ) as chat:
-        for chunk in stream_processor(chat):
-            if chunk.type == "content.delta":
-                aggregated_response += chunk.delta
-            elif chunk.type == "thought.delta":
-                pass
+    if stream:
+        with client.beta.chat.completions.stream(
+            messages=formatted_messages,  # type: ignore
+            model=model_name,
+            temperature=temperature,
+            timeout=httpx.Timeout(30, read=read_timeout),
+            **model_kwargs,
+        ) as chat:
+            for chunk in stream_processor(chat):
+                if chunk.type == "content.delta":
+                    aggregated_response += chunk.delta
+                elif chunk.type == "thought.delta":
+                    pass
+    else:
+        # Non-streaming chat completion
+        chunk = client.beta.chat.completions.parse(
+            messages=formatted_messages,  # type: ignore
+            model=model_name,
+            temperature=temperature,
+            timeout=httpx.Timeout(30, read=read_timeout),
+            **model_kwargs,
+        )
+        aggregated_response = chunk.choices[0].message.content
 
     # Calculate cost of chat
     input_tokens = chunk.usage.prompt_tokens if hasattr(chunk, "usage") and chunk.usage else 0
@@ -182,7 +197,11 @@ async def chat_completion_with_backoff(
         client = get_openai_async_client(openai_api_key, api_base_url)
         openai_async_clients[client_key] = client
 
+    stream = not is_non_streaming_model(model_name, api_base_url)
     stream_processor = adefault_stream_processor
+    if stream:
+        model_kwargs["stream_options"] = {"include_usage": True}
+
     formatted_messages = format_message_for_api(messages, api_base_url)
 
     # Configure thinking for openai reasoning models
@@ -225,12 +244,10 @@ async def chat_completion_with_backoff(
         stream_processor = partial(ain_stream_thought_processor, thought_tag="think")
         # Reasoning is enabled by default. Disable when deepthought is False.
         # See https://qwenlm.github.io/blog/qwen3/#advanced-usages
-        if not deepthought and len(formatted_messages) > 0:
-            formatted_messages[-1]["content"] = formatted_messages[-1]["content"] + " /no_think"
+        if not deepthought:
+            add_qwen_no_think_tag(formatted_messages)
 
-    stream = True
     read_timeout = 300 if is_local_api(api_base_url) else 60
-    model_kwargs["stream_options"] = {"include_usage": True}
     if os.getenv("KHOJ_LLM_SEED"):
         model_kwargs["seed"] = int(os.getenv("KHOJ_LLM_SEED"))
 
@@ -238,7 +255,7 @@ async def chat_completion_with_backoff(
     final_chunk = None
     response_started = False
     start_time = perf_counter()
-    chat_stream: openai.AsyncStream[ChatCompletionChunk] = await client.chat.completions.create(
+    response: openai.AsyncStream[ChatCompletionChunk] | ChatCompletion = await client.chat.completions.create(
         messages=formatted_messages,  # type: ignore
         model=model_name,
         stream=stream,
@@ -246,26 +263,34 @@ async def chat_completion_with_backoff(
         timeout=httpx.Timeout(30, read=read_timeout),
         **model_kwargs,
     )
-    async for chunk in stream_processor(chat_stream):
-        # Log the time taken to start response
-        if not response_started:
-            response_started = True
-            logger.info(f"First response took: {perf_counter() - start_time:.3f} seconds")
-        # Keep track of the last chunk for usage data
-        final_chunk = chunk
-        # Skip empty chunks
-        if len(chunk.choices) == 0:
-            continue
-        # Handle streamed response chunk
-        response_chunk: ResponseWithThought = None
-        response_delta = chunk.choices[0].delta
-        if response_delta.content:
-            response_chunk = ResponseWithThought(response=response_delta.content)
-            aggregated_response += response_chunk.response
-        elif response_delta.thought:
-            response_chunk = ResponseWithThought(thought=response_delta.thought)
-        if response_chunk:
-            yield response_chunk
+    if not stream:
+        # If not streaming, we can return the response directly
+        if len(response.choices) == 0 or not response.choices[0].message:
+            raise ValueError("No response by model.")
+        aggregated_response = response.choices[0].message.content
+        final_chunk = response
+        yield ResponseWithThought(response=aggregated_response)
+    else:
+        async for chunk in stream_processor(response):
+            # Log the time taken to start response
+            if not response_started:
+                response_started = True
+                logger.info(f"First response took: {perf_counter() - start_time:.3f} seconds")
+            # Keep track of the last chunk for usage data
+            final_chunk = chunk
+            # Skip empty chunks
+            if len(chunk.choices) == 0:
+                continue
+            # Handle streamed response chunk
+            response_chunk: ResponseWithThought = None
+            response_delta = chunk.choices[0].delta
+            if response_delta.content:
+                response_chunk = ResponseWithThought(response=response_delta.content)
+                aggregated_response += response_chunk.response
+            elif response_delta.thought:
+                response_chunk = ResponseWithThought(thought=response_delta.thought)
+            if response_chunk:
+                yield response_chunk
 
     # Calculate cost of chat after stream finishes
     input_tokens, output_tokens, cost = 0, 0, 0
@@ -312,11 +337,29 @@ def format_message_for_api(messages: List[ChatMessage], api_base_url: str) -> Li
     """
     formatted_messages = []
     for message in deepcopy(messages):
-        # Convert images to PNG format if message to be sent to non OpenAI API
         if isinstance(message.content, list) and not is_openai_api(api_base_url):
-            for part in message.content:
+            assistant_texts = []
+            has_images = False
+            for idx, part in enumerate(message.content):
+                # Convert images to PNG format if message to be sent to non OpenAI API
                 if part.get("type") == "image_url":
+                    has_images = True
                     part["image_url"]["url"] = convert_image_data_uri(part["image_url"]["url"], target_format="png")
+                # Deepinfra API does not support text content list in assistant messages
+                # So we merge text content list into a single text string
+                if (
+                    part.get("type") == "text"
+                    and message.role == "assistant"
+                    and api_base_url.startswith("https://api.deepinfra.com/v1")
+                ):
+                    assistant_texts += [part["text"]]
+                    message.content.pop(idx)
+            if assistant_texts:
+                assistant_texts_str = "\n\n".join(assistant_texts)
+                if has_images:
+                    message.content += [{"type": "text", "text": assistant_texts_str}]
+                else:
+                    message.content = assistant_texts_str
         formatted_messages.append({"role": message.role, "content": message.content})
 
     return formatted_messages
@@ -336,6 +379,14 @@ def is_openai_reasoning_model(model_name: str, api_base_url: str = None) -> bool
     return model_name.startswith("o") and is_openai_api(api_base_url)
 
 
+def is_non_streaming_model(model_name: str, api_base_url: str = None) -> bool:
+    """
+    Check if model response should not be streamed.
+    """
+    # Some OpenAI models requires biometrics to stream. Avoid streaming their responses.
+    return model_name in ["o3", "o3-pro"] and is_openai_api(api_base_url)
+
+
 def is_twitter_reasoning_model(model_name: str, api_base_url: str = None) -> bool:
     """
     Check if the model is a Twitter reasoning model
@@ -627,3 +678,25 @@ async def ain_stream_thought_processor(
         elif mode == "message":
             chunk.choices[0].delta.content = buf
             yield chunk
+
+
+def add_qwen_no_think_tag(formatted_messages: List[dict]) -> None:
+    """
+    Add /no_think tag to the last message content if it is a user message.
+    This is used to disable reasoning in Qwen models when deepthought is False.
+    """
+    if len(formatted_messages) > 0 and formatted_messages[-1]["role"] == "user":
+        last_message = formatted_messages[-1]
+        if isinstance(last_message["content"], str):
+            # Append /no_think to the last message content
+            formatted_messages[-1]["content"] = last_message["content"] + " /no_think"
+        elif isinstance(last_message["content"], list) and len(last_message["content"]) > 0:
+            # Append /no_think to the last content part
+            if isinstance(last_message["content"][-1], str):
+                last_message["content"][-1] = last_message["content"][-1] + " /no_think"
+            else:
+                # Find last content part of type text and append /no_think to "text" part
+                for content_part in reversed(last_message["content"]):
+                    if isinstance(content_part, dict) and content_part.get("type") == "text":
+                        content_part["text"] += " /no_think"
+                        break

khoj/processor/conversation/utils.py

@@ -62,14 +62,15 @@ model_to_prompt_size = {
     "gpt-4.1": 60000,
     "gpt-4.1-mini": 120000,
     "gpt-4.1-nano": 120000,
-    "o1": 20000,
-    "o3": 30000,
-    "o1-mini": 60000,
-    "o3-mini": 60000,
-    "o4-mini": 60000,
+    "o1-mini": 90000,
+    "o1": 30000,
+    "o3-mini": 90000,
+    "o3": 60000,
+    "o3-pro": 30000,
+    "o4-mini": 90000,
     # Google Models
-    "gemini-2.5-flash-preview-04-17": 120000,
-    "gemini-2.5-pro-preview-03-25": 60000,
+    "gemini-2.5-flash-preview-05-20": 120000,
+    "gemini-2.5-pro-preview-06-05": 60000,
     "gemini-2.0-flash": 120000,
     "gemini-2.0-flash-lite": 120000,
     "gemini-1.5-flash": 120000,
@@ -186,7 +187,7 @@ def construct_iteration_history(
         iteration_history.append(
             ChatMessageModel(
                 by="khoj",
-                intent={"type": "remember", "query": query},
+                intent=Intent(type="remember", query=query),
                 message=previous_iteration_messages,
             )
         )
@@ -196,16 +197,16 @@ def construct_iteration_history(
 def construct_chat_history(chat_history: list[ChatMessageModel], n: int = 4, agent_name="AI") -> str:
     chat_history_str = ""
     for chat in chat_history[-n:]:
-        if chat.by == "khoj" and chat.intent.type in ["remember", "reminder", "summarize"]:
-            if chat.intent.inferred_queries:
-                chat_history_str += f'{agent_name}: {{"queries": {chat.intent.inferred_queries}}}\n'
+        intent_type = chat.intent.type if chat.intent and chat.intent.type else ""
+        inferred_queries = chat.intent.inferred_queries if chat.intent else None
+        if chat.by == "khoj" and intent_type in ["remember", "reminder", "summarize"]:
+            if inferred_queries:
+                chat_history_str += f'{agent_name}: {{"queries": {inferred_queries}}}\n'
             chat_history_str += f"{agent_name}: {chat.message}\n\n"
         elif chat.by == "khoj" and chat.images:
-            chat_history_str += f"User: {chat.intent.query}\n"
             chat_history_str += f"{agent_name}: [generated image redacted for space]\n"
-        elif chat.by == "khoj" and ("excalidraw" in chat.intent.type):
-            chat_history_str += f"User: {chat.intent.query}\n"
-            chat_history_str += f"{agent_name}: {chat.intent.inferred_queries[0]}\n"
+        elif chat.by == "khoj" and ("excalidraw" in intent_type):
+            chat_history_str += f"{agent_name}: {inferred_queries[0]}\n"
         elif chat.by == "you":
             chat_history_str += f"User: {chat.message}\n"
             raw_query_files = chat.queryFiles

khoj/processor/image/generate.py

@@ -53,11 +53,11 @@ async def text_to_image(
     text2image_model = text_to_image_config.model_name
     chat_history_str = ""
     for chat in chat_history[-4:]:
-        if chat.by == "khoj" and chat.intent and chat.intent.type in ["remember", "reminder"]:
-            chat_history_str += f"Q: {chat.intent.query or ''}\n"
+        if chat.by == "you":
+            chat_history_str += f"Q: {chat.message}\n"
+        elif chat.by == "khoj" and chat.intent and chat.intent.type in ["remember", "reminder"]:
             chat_history_str += f"A: {chat.message}\n"
         elif chat.by == "khoj" and chat.images:
-            chat_history_str += f"Q: {chat.intent.query}\n"
             chat_history_str += f"A: Improved Prompt: {chat.intent.inferred_queries[0]}\n"
 
     if send_status_func:

khoj/routers/api_agents.py

@@ -62,7 +62,7 @@ async def all_agents(
     for agent in agents:
         files = agent.fileobject_set.all()
         file_names = [file.file_name for file in files]
-        agent_chat_model = await AgentAdapters.aget_agent_chat_model(default_agent, user)
+        agent_chat_model = await AgentAdapters.aget_agent_chat_model(agent, user)
         agent_packet = {
             "slug": agent.slug,
             "name": agent.name,

khoj/routers/api_chat.py

@@ -960,7 +960,11 @@ async def chat(
             online_results = {key: val.model_dump() for key, val in last_message.onlineContext.items() or []}
             code_results = {key: val.model_dump() for key, val in last_message.codeContext.items() or []}
             compiled_references = [ref.model_dump() for ref in last_message.context or []]
-            research_results = [ResearchIteration(**iter_dict) for iter_dict in last_message.researchContext or []]
+            research_results = [
+                ResearchIteration(**iter_dict)
+                for iter_dict in last_message.researchContext or []
+                if iter_dict.get("summarizedResult")
+            ]
             operator_results = [OperatorRun(**iter_dict) for iter_dict in last_message.operatorContext or []]
             train_of_thought = [thought.model_dump() for thought in last_message.trainOfThought or []]
             # Drop the interrupted message from conversation history
@@ -1011,7 +1015,7 @@ async def chat(
                 user=user,
                 query=defiltered_query,
                 conversation_id=conversation_id,
-                conversation_history=conversation.messages,
+                conversation_history=chat_history,
                 previous_iterations=list(research_results),
                 query_images=uploaded_images,
                 agent=agent,

khoj/utils/constants.py

@@ -17,8 +17,8 @@ default_offline_chat_models = [
     "bartowski/gemma-2-2b-it-GGUF",
     "bartowski/Qwen2.5-14B-Instruct-GGUF",
 ]
-default_openai_chat_models = ["gpt-4o-mini", "gpt-4.1"]
-default_gemini_chat_models = ["gemini-2.0-flash", "gemini-2.5-flash-preview-05-20", "gemini-2.5-pro-preview-05-06"]
+default_openai_chat_models = ["gpt-4o-mini", "gpt-4.1", "o3", "o4-mini"]
+default_gemini_chat_models = ["gemini-2.0-flash", "gemini-2.5-flash-preview-05-20", "gemini-2.5-pro-preview-06-05"]
 default_anthropic_chat_models = ["claude-sonnet-4-0", "claude-3-5-haiku-latest"]
 
 empty_config = {
@@ -41,10 +41,11 @@ model_to_cost: Dict[str, Dict[str, float]] = {
     "gpt-4.1": {"input": 2.00, "output": 8.00},
     "gpt-4.1-mini": {"input": 0.40, "output": 1.60},
     "gpt-4.1-nano": {"input": 0.10, "output": 0.40},
-    "o1": {"input": 15.0, "output": 60.00},
-    "o3": {"input": 10.0, "output": 40.00},
     "o1-mini": {"input": 3.0, "output": 12.0},
+    "o1": {"input": 15.0, "output": 60.00},
     "o3-mini": {"input": 1.10, "output": 4.40},
+    "o3": {"input": 2.0, "output": 8.00},
+    "o3-pro": {"input": 20.0, "output": 80.00},
     "o4-mini": {"input": 1.10, "output": 4.40},
     # Gemini Pricing: https://ai.google.dev/pricing
     "gemini-1.5-flash": {"input": 0.075, "output": 0.30},
@@ -53,8 +54,8 @@ model_to_cost: Dict[str, Dict[str, float]] = {
     "gemini-1.5-pro-002": {"input": 1.25, "output": 5.00},
     "gemini-2.0-flash": {"input": 0.10, "output": 0.40},
     "gemini-2.0-flash-lite": {"input": 0.0075, "output": 0.30},
-    "gemini-2.5-flash-preview-04-17": {"input": 0.15, "output": 0.60, "thought": 3.50},
-    "gemini-2.5-pro-preview-03-25": {"input": 1.25, "output": 10.0},
+    "gemini-2.5-flash-preview-05-20": {"input": 0.15, "output": 0.60, "thought": 3.50},
+    "gemini-2.5-pro-preview-06-05": {"input": 1.25, "output": 10.0},
     # Anthropic Pricing: https://www.anthropic.com/pricing#anthropic-api
     "claude-3-5-haiku-20241022": {"input": 1.0, "output": 5.0, "cache_read": 0.08, "cache_write": 1.0},
     "claude-3-5-haiku@20241022": {"input": 1.0, "output": 5.0, "cache_read": 0.08, "cache_write": 1.0},

{khoj-1.42.3.dev1.dist-info → khoj-1.42.4.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: khoj
-Version: 1.42.3.dev1
+Version: 1.42.4
 Summary: Your Second Brain
 Project-URL: Homepage, https://khoj.dev
 Project-URL: Documentation, https://docs.khoj.dev
@@ -53,7 +53,7 @@ Requires-Dist: magika~=0.5.1
 Requires-Dist: markdown-it-py~=3.0.0
 Requires-Dist: markdownify~=0.11.6
 Requires-Dist: openai-whisper>=20231117
-Requires-Dist: openai>=1.0.0
+Requires-Dist: openai>=1.86.0
 Requires-Dist: pgvector==0.2.4
 Requires-Dist: phonenumbers==8.13.27
 Requires-Dist: pillow~=10.0.0