Improve model token limit detection (#3292)

* Properly find context window for ollama llama

* Better ollama support + upgrade litellm

* Ugprade OpenAI as well

* Fix mypy

backend/danswer/configs/model_configs.py

@@ -70,7 +70,9 @@ GEN_AI_NUM_RESERVED_OUTPUT_TOKENS = int(
 )
 
 # Typically, GenAI models nowadays are at least 4K tokens
-GEN_AI_MODEL_FALLBACK_MAX_TOKENS = 4096
+GEN_AI_MODEL_FALLBACK_MAX_TOKENS = int(
+    os.environ.get("GEN_AI_MODEL_FALLBACK_MAX_TOKENS") or 4096
+)
 
 # Number of tokens from chat history to include at maximum
 # 3000 should be enough context regardless of use, no need to include as much as possible

backend/danswer/llm/chat_llm.py

@@ -26,7 +26,9 @@ from langchain_core.messages.tool import ToolMessage
 from langchain_core.prompt_values import PromptValue
 
 from danswer.configs.app_configs import LOG_DANSWER_MODEL_INTERACTIONS
-from danswer.configs.model_configs import DISABLE_LITELLM_STREAMING
+from danswer.configs.model_configs import (
+    DISABLE_LITELLM_STREAMING,
+)
 from danswer.configs.model_configs import GEN_AI_TEMPERATURE
 from danswer.configs.model_configs import LITELLM_EXTRA_BODY
 from danswer.llm.interfaces import LLM
@@ -161,7 +163,9 @@ def _convert_delta_to_message_chunk(
 
     if role == "user":
         return HumanMessageChunk(content=content)
-    elif role == "assistant":
+    # NOTE: if tool calls are present, then it's an assistant.
+    # In Ollama, the role will be None for tool-calls
+    elif role == "assistant" or tool_calls:
         if tool_calls:
             tool_call = tool_calls[0]
             tool_name = tool_call.function.name or (curr_msg and curr_msg.name) or ""
@@ -236,6 +240,7 @@ class DefaultMultiLLM(LLM):
         custom_config: dict[str, str] | None = None,
         extra_headers: dict[str, str] | None = None,
         extra_body: dict | None = LITELLM_EXTRA_BODY,
+        model_kwargs: dict[str, Any] | None = None,
         long_term_logger: LongTermLogger | None = None,
     ):
         self._timeout = timeout
@@ -268,7 +273,7 @@ class DefaultMultiLLM(LLM):
             for k, v in custom_config.items():
                 os.environ[k] = v
 
-        model_kwargs: dict[str, Any] = {}
+        model_kwargs = model_kwargs or {}
         if extra_headers:
             model_kwargs.update({"extra_headers": extra_headers})
         if extra_body:

backend/danswer/llm/factory.py

@@ -1,5 +1,8 @@
+from typing import Any
+
 from danswer.configs.app_configs import DISABLE_GENERATIVE_AI
 from danswer.configs.chat_configs import QA_TIMEOUT
+from danswer.configs.model_configs import GEN_AI_MODEL_FALLBACK_MAX_TOKENS
 from danswer.configs.model_configs import GEN_AI_TEMPERATURE
 from danswer.db.engine import get_session_context_manager
 from danswer.db.llm import fetch_default_provider
@@ -13,6 +16,15 @@ from danswer.utils.headers import build_llm_extra_headers
 from danswer.utils.long_term_log import LongTermLogger
 
 
+def _build_extra_model_kwargs(provider: str) -> dict[str, Any]:
+    """Ollama requires us to specify the max context window.
+
+    For now, just using the GEN_AI_MODEL_FALLBACK_MAX_TOKENS value.
+    TODO: allow model-specific values to be configured via the UI.
+    """
+    return {"num_ctx": GEN_AI_MODEL_FALLBACK_MAX_TOKENS} if provider == "ollama" else {}
+
+
 def get_main_llm_from_tuple(
     llms: tuple[LLM, LLM],
 ) -> LLM:
@@ -132,5 +144,6 @@ def get_llm(
         temperature=temperature,
         custom_config=custom_config,
         extra_headers=build_llm_extra_headers(additional_headers),
+        model_kwargs=_build_extra_model_kwargs(provider),
         long_term_logger=long_term_logger,
     )

backend/danswer/llm/utils.py

@@ -1,3 +1,4 @@
+import copy
 import io
 import json
 from collections.abc import Callable
@@ -385,6 +386,62 @@ def test_llm(llm: LLM) -> str | None:
     return error_msg
 
 
+def get_model_map() -> dict:
+    starting_map = copy.deepcopy(cast(dict, litellm.model_cost))
+
+    # NOTE: we could add additional models here in the future,
+    # but for now there is no point. Ollama allows the user to
+    # to specify their desired max context window, and it's
+    # unlikely to be standard across users even for the same model
+    # (it heavily depends on their hardware). For now, we'll just
+    # rely on GEN_AI_MODEL_FALLBACK_MAX_TOKENS to cover this.
+    # for model_name in [
+    #     "llama3.2",
+    #     "llama3.2:1b",
+    #     "llama3.2:3b",
+    #     "llama3.2:11b",
+    #     "llama3.2:90b",
+    # ]:
+    #     starting_map[f"ollama/{model_name}"] = {
+    #         "max_tokens": 128000,
+    #         "max_input_tokens": 128000,
+    #         "max_output_tokens": 128000,
+    #     }
+
+    return starting_map
+
+
+def _strip_extra_provider_from_model_name(model_name: str) -> str:
+    return model_name.split("/")[1] if "/" in model_name else model_name
+
+
+def _strip_colon_from_model_name(model_name: str) -> str:
+    return ":".join(model_name.split(":")[:-1]) if ":" in model_name else model_name
+
+
+def _find_model_obj(
+    model_map: dict, provider: str, model_names: list[str | None]
+) -> dict | None:
+    # Filter out None values and deduplicate model names
+    filtered_model_names = [name for name in model_names if name]
+
+    # First try all model names with provider prefix
+    for model_name in filtered_model_names:
+        model_obj = model_map.get(f"{provider}/{model_name}")
+        if model_obj:
+            logger.debug(f"Using model object for {provider}/{model_name}")
+            return model_obj
+
+    # Then try all model names without provider prefix
+    for model_name in filtered_model_names:
+        model_obj = model_map.get(model_name)
+        if model_obj:
+            logger.debug(f"Using model object for {model_name}")
+            return model_obj
+
+    return None
+
+
 def get_llm_max_tokens(
     model_map: dict,
     model_name: str,
@@ -397,22 +454,22 @@ def get_llm_max_tokens(
         return GEN_AI_MAX_TOKENS
 
     try:
-        model_obj = model_map.get(f"{model_provider}/{model_name}")
+        extra_provider_stripped_model_name = _strip_extra_provider_from_model_name(
-        if model_obj:
+            model_name
-            logger.debug(f"Using model object for {model_provider}/{model_name}")
+        )
-
+        model_obj = _find_model_obj(
-        if not model_obj:
+            model_map,
-            model_obj = model_map.get(model_name)
+            model_provider,
-            if model_obj:
+            [
-                logger.debug(f"Using model object for {model_name}")
+                model_name,
-
+                # Remove leading extra provider. Usually for cases where user has a
-        if not model_obj:
+                # customer model proxy which appends another prefix
-            model_name_split = model_name.split("/")
+                extra_provider_stripped_model_name,
-            if len(model_name_split) > 1:
+                # remove :XXXX from the end, if present. Needed for ollama.
-                model_obj = model_map.get(model_name_split[1])
+                _strip_colon_from_model_name(model_name),
-            if model_obj:
+                _strip_colon_from_model_name(extra_provider_stripped_model_name),
-                logger.debug(f"Using model object for {model_name_split[1]}")
+            ],
-
+        )
         if not model_obj:
             raise RuntimeError(
                 f"No litellm entry found for {model_provider}/{model_name}"
@@ -488,7 +545,7 @@ def get_max_input_tokens(
     # `model_cost` dict is a named public interface:
     # https://litellm.vercel.app/docs/completion/token_usage#7-model_cost
     # model_map is  litellm.model_cost
-    litellm_model_map = litellm.model_cost
+    litellm_model_map = get_model_map()
 
     input_toks = (
         get_llm_max_tokens(

backend/requirements/default.txt

@@ -29,7 +29,7 @@ trafilatura==1.12.2
 langchain==0.1.17
 langchain-core==0.1.50
 langchain-text-splitters==0.0.1
-litellm==1.50.2
+litellm==1.53.1
 lxml==5.3.0
 lxml_html_clean==0.2.2
 llama-index==0.9.45
@@ -38,7 +38,7 @@ msal==1.28.0
 nltk==3.8.1
 Office365-REST-Python-Client==2.5.9
 oauthlib==3.2.2
-openai==1.52.2
+openai==1.55.3
 openpyxl==3.1.2
 playwright==1.41.2
 psutil==5.9.5