Support Claude 3.7 and use its extended thinking in research mode

Claude 3.7 Sonnet is Anthropic's first reasoning model. It provides a
single model/api capable of standard and extended thinking. Utilize
the extended thinking in Khoj's research mode.

Increase default max output tokens to 8K for Anthropic models.

pyproject.toml

@@ -86,7 +86,7 @@ dependencies = [
     "pytz ~= 2024.1",
     "cron-descriptor == 1.4.3",
     "django_apscheduler == 0.6.2",
-    "anthropic == 0.26.1",
+    "anthropic == 0.49.0",
     "docx2txt == 0.8",
     "google-generativeai == 0.8.3",
     "pyjson5 == 1.6.7",

src/khoj/processor/conversation/anthropic/anthropic_chat.py

@@ -31,7 +31,7 @@ logger = logging.getLogger(__name__)
 
 def extract_questions_anthropic(
     text,
-    model: Optional[str] = "claude-instant-1.2",
+    model: Optional[str] = "claude-3-7-sonnet-latest",
     conversation_log={},
     api_key=None,
     temperature=0.7,
@@ -122,7 +122,7 @@ def extract_questions_anthropic(
     return questions
 
 
-def anthropic_send_message_to_model(messages, api_key, model, response_type="text", tracer={}):
+def anthropic_send_message_to_model(messages, api_key, model, response_type="text", deepthought=False, tracer={}):
     """
     Send message to model
     """
@@ -135,6 +135,7 @@ def anthropic_send_message_to_model(messages, api_key, model, response_type="tex
         model_name=model,
         api_key=api_key,
         response_type=response_type,
+        deepthought=deepthought,
         tracer=tracer,
     )
 
@@ -145,7 +146,7 @@ def converse_anthropic(
     online_results: Optional[Dict[str, Dict]] = None,
     code_results: Optional[Dict[str, Dict]] = None,
     conversation_log={},
-    model: Optional[str] = "claude-3-5-sonnet-20241022",
+    model: Optional[str] = "claude-3-7-sonnet-latest",
     api_key: Optional[str] = None,
     completion_func=None,
     conversation_commands=[ConversationCommand.Default],
@@ -160,6 +161,7 @@ def converse_anthropic(
     generated_files: List[FileAttachment] = None,
     program_execution_context: Optional[List[str]] = None,
     generated_asset_results: Dict[str, Dict] = {},
+    deepthought: Optional[bool] = False,
     tracer: dict = {},
 ):
     """
@@ -239,5 +241,6 @@ def converse_anthropic(
         system_prompt=system_prompt,
         completion_func=completion_func,
         max_prompt_size=max_prompt_size,
+        deepthought=deepthought,
         tracer=tracer,
     )

src/khoj/processor/conversation/anthropic/utils.py

@@ -17,10 +17,8 @@ from khoj.processor.conversation.utils import (
     commit_conversation_trace,
     get_image_from_url,
 )
-from khoj.utils import state
 from khoj.utils.helpers import (
     get_chat_usage_metrics,
-    in_debug_mode,
     is_none_or_empty,
     is_promptrace_enabled,
 )
@@ -30,7 +28,8 @@ logger = logging.getLogger(__name__)
 anthropic_clients: Dict[str, anthropic.Anthropic] = {}
 
 
-DEFAULT_MAX_TOKENS_ANTHROPIC = 3000
+DEFAULT_MAX_TOKENS_ANTHROPIC = 8000
+MAX_REASONING_TOKENS_ANTHROPIC = 12000
 
 
 @retry(
@@ -42,12 +41,13 @@ DEFAULT_MAX_TOKENS_ANTHROPIC = 3000
 def anthropic_completion_with_backoff(
     messages,
     system_prompt,
-    model_name,
+    model_name: str,
     temperature=0,
     api_key=None,
     model_kwargs=None,
     max_tokens=None,
     response_type="text",
+    deepthought=False,
     tracer={},
 ) -> str:
     if api_key not in anthropic_clients:
@@ -57,18 +57,24 @@ def anthropic_completion_with_backoff(
         client = anthropic_clients[api_key]
 
     formatted_messages = [{"role": message.role, "content": message.content} for message in messages]
-    if response_type == "json_object":
+    aggregated_response = ""
-        # Prefill model response with '{' to make it output a valid JSON object
+    if response_type == "json_object" and not deepthought:
+        # Prefill model response with '{' to make it output a valid JSON object. Not supported with extended thinking.
         formatted_messages += [{"role": "assistant", "content": "{"}]
-
+        aggregated_response += "{"
-    aggregated_response = "{" if response_type == "json_object" else ""
-    max_tokens = max_tokens or DEFAULT_MAX_TOKENS_ANTHROPIC
 
     final_message = None
     model_kwargs = model_kwargs or dict()
     if system_prompt:
         model_kwargs["system"] = system_prompt
 
+    max_tokens = max_tokens or DEFAULT_MAX_TOKENS_ANTHROPIC
+    if deepthought and model_name.startswith("claude-3-7"):
+        model_kwargs["thinking"] = {"type": "enabled", "budget_tokens": MAX_REASONING_TOKENS_ANTHROPIC}
+        max_tokens += MAX_REASONING_TOKENS_ANTHROPIC
+        # Temperature control not supported when using extended thinking
+        temperature = 1.0
+
     with client.messages.stream(
         messages=formatted_messages,
         model=model_name,  # type: ignore
@@ -111,20 +117,41 @@ def anthropic_chat_completion_with_backoff(
     system_prompt,
     max_prompt_size=None,
     completion_func=None,
+    deepthought=False,
     model_kwargs=None,
     tracer={},
 ):
     g = ThreadedGenerator(compiled_references, online_results, completion_func=completion_func)
     t = Thread(
         target=anthropic_llm_thread,
-        args=(g, messages, system_prompt, model_name, temperature, api_key, max_prompt_size, model_kwargs, tracer),
+        args=(
+            g,
+            messages,
+            system_prompt,
+            model_name,
+            temperature,
+            api_key,
+            max_prompt_size,
+            deepthought,
+            model_kwargs,
+            tracer,
+        ),
     )
     t.start()
     return g
 
 
 def anthropic_llm_thread(
-    g, messages, system_prompt, model_name, temperature, api_key, max_prompt_size=None, model_kwargs=None, tracer={}
+    g,
+    messages,
+    system_prompt,
+    model_name,
+    temperature,
+    api_key,
+    max_prompt_size=None,
+    deepthought=False,
+    model_kwargs=None,
+    tracer={},
 ):
     try:
         if api_key not in anthropic_clients:
@@ -133,6 +160,14 @@ def anthropic_llm_thread(
         else:
             client: anthropic.Anthropic = anthropic_clients[api_key]
 
+        model_kwargs = model_kwargs or dict()
+        max_tokens = DEFAULT_MAX_TOKENS_ANTHROPIC
+        if deepthought and model_name.startswith("claude-3-7"):
+            model_kwargs["thinking"] = {"type": "enabled", "budget_tokens": MAX_REASONING_TOKENS_ANTHROPIC}
+            max_tokens += MAX_REASONING_TOKENS_ANTHROPIC
+            # Temperature control not supported when using extended thinking
+            temperature = 1.0
+
         formatted_messages: List[anthropic.types.MessageParam] = [
             anthropic.types.MessageParam(role=message.role, content=message.content) for message in messages
         ]
@@ -145,8 +180,8 @@ def anthropic_llm_thread(
             temperature=temperature,
             system=system_prompt,
             timeout=20,
-            max_tokens=DEFAULT_MAX_TOKENS_ANTHROPIC,
+            max_tokens=max_tokens,
-            **(model_kwargs or dict()),
+            **model_kwargs,
         ) as stream:
             for text in stream.text_stream:
                 aggregated_response += text

src/khoj/processor/conversation/utils.py

@@ -61,6 +61,9 @@ model_to_prompt_size = {
     "gemini-1.5-pro": 60000,
     # Anthropic Models
     "claude-3-5-sonnet-20241022": 60000,
+    "claude-3-5-sonnet-latest": 60000,
+    "claude-3-7-sonnet-20250219": 60000,
+    "claude-3-7-sonnet-latest": 60000,
     "claude-3-5-haiku-20241022": 60000,
     # Offline Models
     "bartowski/Qwen2.5-14B-Instruct-GGUF": 20000,

src/khoj/routers/helpers.py

@@ -1125,6 +1125,7 @@ async def send_message_to_model_wrapper(
     query: str,
     system_message: str = "",
     response_type: str = "text",
+    deepthought: bool = False,
     user: KhojUser = None,
     query_images: List[str] = None,
     context: str = "",
@@ -1227,6 +1228,7 @@ async def send_message_to_model_wrapper(
             api_key=api_key,
             model=chat_model_name,
             response_type=response_type,
+            deepthought=deepthought,
             tracer=tracer,
         )
     elif model_type == ChatModel.ModelType.GOOGLE:
@@ -1425,11 +1427,13 @@ def generate_chat_response(
         )
 
         query_to_run = q
+        deepthought = False
         if meta_research:
             query_to_run = f"<query>{q}</query>\n<collected_research>\n{meta_research}\n</collected_research>"
             compiled_references = []
             online_results = {}
             code_results = {}
+            deepthought = True
 
         chat_model = ConversationAdapters.get_valid_chat_model(user, conversation, is_subscribed)
         vision_available = chat_model.vision_enabled
@@ -1513,6 +1517,7 @@ def generate_chat_response(
                 generated_files=raw_generated_files,
                 generated_asset_results=generated_asset_results,
                 program_execution_context=program_execution_context,
+                deepthought=deepthought,
                 tracer=tracer,
             )
         elif chat_model.model_type == ChatModel.ModelType.GOOGLE:

src/khoj/routers/research.py

@@ -95,6 +95,7 @@ async def apick_next_tool(
                 query=query,
                 context=function_planning_prompt,
                 response_type="json_object",
+                deepthought=True,
                 user=user,
                 query_images=query_images,
                 query_files=query_files,

src/khoj/utils/constants.py

@@ -48,6 +48,9 @@ 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},
     # Anthropic Pricing: https://www.anthropic.com/pricing#anthropic-api_
-    "claude-3-5-sonnet-20241022": {"input": 3.0, "output": 15.0},
     "claude-3-5-haiku-20241022": {"input": 1.0, "output": 5.0},
+    "claude-3-5-sonnet-20241022": {"input": 3.0, "output": 15.0},
+    "claude-3-5-sonnet-latest": {"input": 3.0, "output": 15.0},
+    "claude-3-7-sonnet-20250219": {"input": 3.0, "output": 15.0},
+    "claude-3-7-sonnet-latest": {"input": 3.0, "output": 15.0},
 }