Show details for query tokens (#1392)

embedchain/config/llm/base.py

@@ -1,3 +1,4 @@
+import json
 import logging
 import re
 from string import Template
@@ -92,6 +93,7 @@ class BaseLlmConfig(BaseConfig):
         top_p: float = 1,
         stream: bool = False,
         online: bool = False,
+        token_usage: bool = False,
         deployment_name: Optional[str] = None,
         system_prompt: Optional[str] = None,
         where: dict[str, Any] = None,
@@ -135,6 +137,8 @@ class BaseLlmConfig(BaseConfig):
         :type stream: bool, optional
         :param online: Controls whether to use internet for answering query, defaults to False
         :type online: bool, optional
+        :param token_usage: Controls whether to return token usage in response, defaults to False
+        :type token_usage: bool, optional
         :param deployment_name: t.b.a., defaults to None
         :type deployment_name: Optional[str], optional
         :param system_prompt: System prompt string, defaults to None
@@ -180,6 +184,8 @@ class BaseLlmConfig(BaseConfig):
         self.max_tokens = max_tokens
         self.model = model
         self.top_p = top_p
+        self.online = online
+        self.token_usage = token_usage
         self.deployment_name = deployment_name
         self.system_prompt = system_prompt
         self.query_type = query_type
@@ -197,6 +203,10 @@ class BaseLlmConfig(BaseConfig):
         self.online = online
         self.api_version = api_version
 
+        if token_usage:
+            f = open("model_prices_and_context_window.json")
+            self.model_pricing_map = json.load(f)
+
         if isinstance(prompt, str):
             prompt = Template(prompt)
 

embedchain/embedchain.py

@@ -6,9 +6,7 @@ from typing import Any, Optional, Union
 from dotenv import load_dotenv
 from langchain.docstore.document import Document
 
-from embedchain.cache import (adapt, get_gptcache_session,
-                              gptcache_data_convert,
-                              gptcache_update_cache_callback)
+from embedchain.cache import adapt, get_gptcache_session, gptcache_data_convert, gptcache_update_cache_callback
 from embedchain.chunkers.base_chunker import BaseChunker
 from embedchain.config import AddConfig, BaseLlmConfig, ChunkerConfig
 from embedchain.config.base_app_config import BaseAppConfig
@@ -18,8 +16,7 @@ from embedchain.embedder.base import BaseEmbedder
 from embedchain.helpers.json_serializable import JSONSerializable
 from embedchain.llm.base import BaseLlm
 from embedchain.loaders.base_loader import BaseLoader
-from embedchain.models.data_type import (DataType, DirectDataType,
-                                         IndirectDataType, SpecialDataType)
+from embedchain.models.data_type import DataType, DirectDataType, IndirectDataType, SpecialDataType
 from embedchain.utils.misc import detect_datatype, is_valid_json_string
 from embedchain.vectordb.base import BaseVectorDB
 
@@ -478,7 +475,7 @@ class EmbedChain(JSONSerializable):
         where: Optional[dict] = None,
         citations: bool = False,
         **kwargs: dict[str, Any],
-    ) -> Union[tuple[str, list[tuple[str, dict]]], str]:
+    ) -> Union[tuple[str, list[tuple[str, dict]]], str, dict[str, Any]]:
         """
         Queries the vector database based on the given input query.
         Gets relevant doc based on the query and then passes it to an
@@ -501,7 +498,9 @@ class EmbedChain(JSONSerializable):
         :type kwargs: dict[str, Any]
         :return: The answer to the query, with citations if the citation flag is True
         or the dry run result
-        :rtype: str, if citations is False, otherwise tuple[str, list[tuple[str,str,str]]]
+        :rtype: str, if citations is False and token_usage is False, otherwise if citations is true then
+        tuple[str, list[tuple[str,str,str]]] and if token_usage is true then
+        tuple[str, list[tuple[str,str,str]], dict[str, Any]]
         """
         contexts = self._retrieve_from_database(
             input_query=input_query, config=config, where=where, citations=citations, **kwargs
@@ -524,17 +523,29 @@ class EmbedChain(JSONSerializable):
                 dry_run=dry_run,
             )
         else:
-            answer = self.llm.query(
-                input_query=input_query, contexts=contexts_data_for_llm_query, config=config, dry_run=dry_run
-            )
+            if self.llm.config.token_usage:
+                answer, token_info = self.llm.query(
+                    input_query=input_query, contexts=contexts_data_for_llm_query, config=config, dry_run=dry_run
+                )
+            else:
+                answer = self.llm.query(
+                    input_query=input_query, contexts=contexts_data_for_llm_query, config=config, dry_run=dry_run
+                )
 
         # Send anonymous telemetry
         self.telemetry.capture(event_name="query", properties=self._telemetry_props)
 
         if citations:
+            if self.llm.config.token_usage:
+                return {"answer": answer, "contexts": contexts, "usage": token_info}
             return answer, contexts
-        else:
-            return answer
+        if self.llm.config.token_usage:
+            return {"answer": answer, "usage": token_info}
+
+        logger.warning(
+            "Starting from v0.1.125 the return type of query method will be changed to tuple containing `answer`."
+        )
+        return answer
 
     def chat(
         self,
@@ -545,7 +556,7 @@ class EmbedChain(JSONSerializable):
         where: Optional[dict[str, str]] = None,
         citations: bool = False,
         **kwargs: dict[str, Any],
-    ) -> Union[tuple[str, list[tuple[str, dict]]], str]:
+    ) -> Union[tuple[str, list[tuple[str, dict]]], str, dict[str, Any]]:
         """
         Queries the vector database on the given input query.
         Gets relevant doc based on the query and then passes it to an
@@ -572,7 +583,9 @@ class EmbedChain(JSONSerializable):
         :type kwargs: dict[str, Any]
         :return: The answer to the query, with citations if the citation flag is True
         or the dry run result
-        :rtype: str, if citations is False, otherwise tuple[str, list[tuple[str,str,str]]]
+        :rtype: str, if citations is False and token_usage is False, otherwise if citations is true then
+        tuple[str, list[tuple[str,str,str]]] and if token_usage is true then
+        tuple[str, list[tuple[str,str,str]], dict[str, Any]]
         """
         contexts = self._retrieve_from_database(
             input_query=input_query, config=config, where=where, citations=citations, **kwargs
@@ -600,9 +613,14 @@ class EmbedChain(JSONSerializable):
             )
         else:
             logger.debug("Cache disabled. Running chat without cache.")
-            answer = self.llm.chat(
-                input_query=input_query, contexts=contexts_data_for_llm_query, config=config, dry_run=dry_run
-            )
+            if self.llm.config.token_usage:
+                answer, token_info = self.llm.query(
+                    input_query=input_query, contexts=contexts_data_for_llm_query, config=config, dry_run=dry_run
+                )
+            else:
+                answer = self.llm.query(
+                    input_query=input_query, contexts=contexts_data_for_llm_query, config=config, dry_run=dry_run
+                )
 
         # add conversation in memory
         self.llm.add_history(self.config.id, input_query, answer, session_id=session_id)
@@ -611,9 +629,16 @@ class EmbedChain(JSONSerializable):
         self.telemetry.capture(event_name="chat", properties=self._telemetry_props)
 
         if citations:
+            if self.llm.config.token_usage:
+                return {"answer": answer, "contexts": contexts, "usage": token_info}
             return answer, contexts
-        else:
-            return answer
+        if self.llm.config.token_usage:
+            return {"answer": answer, "usage": token_info}
+
+        logger.warning(
+            "Starting from v0.1.125 the return type of query method will be changed to tuple containing `answer`."
+        )
+        return answer
 
     def search(self, query, num_documents=3, where=None, raw_filter=None, namespace=None):
         """

embedchain/embedder/gpt4all.py

@@ -9,10 +9,10 @@ class GPT4AllEmbedder(BaseEmbedder):
     def __init__(self, config: Optional[BaseEmbedderConfig] = None):
         super().__init__(config=config)
 
-        from langchain.embeddings import \
-            GPT4AllEmbeddings as LangchainGPT4AllEmbeddings
+        from langchain_community.embeddings import GPT4AllEmbeddings as LangchainGPT4AllEmbeddings
 
-        embeddings = LangchainGPT4AllEmbeddings()
+        model_name = self.config.model or "all-MiniLM-L6-v2-f16.gguf"
+        embeddings = LangchainGPT4AllEmbeddings(model_name=model_name)
         embedding_fn = BaseEmbedder._langchain_default_concept(embeddings)
         self.set_embedding_fn(embedding_fn=embedding_fn)
 

embedchain/llm/anthropic.py

@@ -1,6 +1,6 @@
 import logging
 import os
-from typing import Optional
+from typing import Any, Optional
 
 try:
     from langchain_anthropic import ChatAnthropic
@@ -21,8 +21,27 @@ class AnthropicLlm(BaseLlm):
         if not self.config.api_key and "ANTHROPIC_API_KEY" not in os.environ:
             raise ValueError("Please set the ANTHROPIC_API_KEY environment variable or pass it in the config.")
 
-    def get_llm_model_answer(self, prompt):
-        return AnthropicLlm._get_answer(prompt=prompt, config=self.config)
+    def get_llm_model_answer(self, prompt) -> tuple[str, Optional[dict[str, Any]]]:
+        if self.config.token_usage:
+            response, token_info = self._get_answer(prompt, self.config)
+            model_name = "anthropic/" + self.config.model
+            if model_name not in self.config.model_pricing_map:
+                raise ValueError(
+                    f"Model {model_name} not found in `model_prices_and_context_window.json`. \
+                    You can disable token usage by setting `token_usage` to False."
+                )
+            total_cost = (
+                self.config.model_pricing_map[model_name]["input_cost_per_token"] * token_info["input_tokens"]
+            ) + self.config.model_pricing_map[model_name]["output_cost_per_token"] * token_info["output_tokens"]
+            response_token_info = {
+                "prompt_tokens": token_info["input_tokens"],
+                "completion_tokens": token_info["output_tokens"],
+                "total_tokens": token_info["input_tokens"] + token_info["output_tokens"],
+                "total_cost": round(total_cost, 10),
+                "cost_currency": "USD",
+            }
+            return response, response_token_info
+        return self._get_answer(prompt, self.config)
 
     @staticmethod
     def _get_answer(prompt: str, config: BaseLlmConfig) -> str:
@@ -34,4 +53,7 @@ class AnthropicLlm(BaseLlm):
 
         messages = BaseLlm._get_messages(prompt, system_prompt=config.system_prompt)
 
-        return chat(messages).content
+        chat_response = chat.invoke(messages)
+        if config.token_usage:
+            return chat_response.content, chat_response.response_metadata["token_usage"]
+        return chat_response.content

embedchain/llm/base.py

@@ -164,7 +164,7 @@ class BaseLlm(JSONSerializable):
         return search.run(input_query)
 
     @staticmethod
-    def _stream_response(answer: Any) -> Generator[Any, Any, None]:
+    def _stream_response(answer: Any, token_info: Optional[dict[str, Any]] = None) -> Generator[Any, Any, None]:
         """Generator to be used as streaming response
 
         :param answer: Answer chunk from llm
@@ -177,6 +177,8 @@ class BaseLlm(JSONSerializable):
             streamed_answer = streamed_answer + chunk
             yield chunk
         logger.info(f"Answer: {streamed_answer}")
+        if token_info:
+            logger.info(f"Token Info: {token_info}")
 
     def query(self, input_query: str, contexts: list[str], config: BaseLlmConfig = None, dry_run=False):
         """
@@ -219,11 +221,18 @@ class BaseLlm(JSONSerializable):
             if dry_run:
                 return prompt
 
-            answer = self.get_answer_from_llm(prompt)
+            if self.config.token_usage:
+                answer, token_info = self.get_answer_from_llm(prompt)
+            else:
+                answer = self.get_answer_from_llm(prompt)
             if isinstance(answer, str):
                 logger.info(f"Answer: {answer}")
+                if self.config.token_usage:
+                    return answer, token_info
                 return answer
             else:
+                if self.config.token_usage:
+                    return self._stream_response(answer, token_info)
                 return self._stream_response(answer)
         finally:
             if config:
@@ -276,13 +285,13 @@ class BaseLlm(JSONSerializable):
             if dry_run:
                 return prompt
 
-            answer = self.get_answer_from_llm(prompt)
+            answer, token_info = self.get_answer_from_llm(prompt)
             if isinstance(answer, str):
                 logger.info(f"Answer: {answer}")
-                return answer
+                return answer, token_info
             else:
                 # this is a streamed response and needs to be handled differently.
-                return self._stream_response(answer)
+                return self._stream_response(answer, token_info)
         finally:
             if config:
                 # Restore previous config

embedchain/llm/cohere.py

@@ -1,8 +1,8 @@
 import importlib
 import os
-from typing import Optional
+from typing import Any, Optional
 
-from langchain_community.llms.cohere import Cohere
+from langchain_cohere import ChatCohere
 
 from embedchain.config import BaseLlmConfig
 from embedchain.helpers.json_serializable import register_deserializable
@@ -17,27 +17,50 @@ class CohereLlm(BaseLlm):
         except ModuleNotFoundError:
             raise ModuleNotFoundError(
                 "The required dependencies for Cohere are not installed."
-                'Please install with `pip install --upgrade "embedchain[cohere]"`'
+                "Please install with `pip install langchain_cohere==1.16.0`"
             ) from None
 
         super().__init__(config=config)
         if not self.config.api_key and "COHERE_API_KEY" not in os.environ:
             raise ValueError("Please set the COHERE_API_KEY environment variable or pass it in the config.")
 
-    def get_llm_model_answer(self, prompt):
+    def get_llm_model_answer(self, prompt) -> tuple[str, Optional[dict[str, Any]]]:
         if self.config.system_prompt:
             raise ValueError("CohereLlm does not support `system_prompt`")
-        return CohereLlm._get_answer(prompt=prompt, config=self.config)
+
+        if self.config.token_usage:
+            response, token_info = self._get_answer(prompt, self.config)
+            model_name = "cohere/" + self.config.model
+            if model_name not in self.config.model_pricing_map:
+                raise ValueError(
+                    f"Model {model_name} not found in `model_prices_and_context_window.json`. \
+                    You can disable token usage by setting `token_usage` to False."
+                )
+            total_cost = (
+                self.config.model_pricing_map[model_name]["input_cost_per_token"] * token_info["input_tokens"]
+            ) + self.config.model_pricing_map[model_name]["output_cost_per_token"] * token_info["output_tokens"]
+            response_token_info = {
+                "prompt_tokens": token_info["input_tokens"],
+                "completion_tokens": token_info["output_tokens"],
+                "total_tokens": token_info["input_tokens"] + token_info["output_tokens"],
+                "total_cost": round(total_cost, 10),
+                "cost_currency": "USD",
+            }
+            return response, response_token_info
+        return self._get_answer(prompt, self.config)
 
     @staticmethod
     def _get_answer(prompt: str, config: BaseLlmConfig) -> str:
-        api_key = config.api_key or os.getenv("COHERE_API_KEY")
-        llm = Cohere(
-            cohere_api_key=api_key,
-            model=config.model,
-            max_tokens=config.max_tokens,
-            temperature=config.temperature,
-            p=config.top_p,
-        )
+        api_key = config.api_key or os.environ["COHERE_API_KEY"]
+        kwargs = {
+            "model_name": config.model or "command-r",
+            "temperature": config.temperature,
+            "max_tokens": config.max_tokens,
+            "together_api_key": api_key,
+        }
 
-        return llm.invoke(prompt)
+        chat = ChatCohere(**kwargs)
+        chat_response = chat.invoke(prompt)
+        if config.token_usage:
+            return chat_response.content, chat_response.response_metadata["token_count"]
+        return chat_response.content

embedchain/llm/groq.py

@@ -1,5 +1,5 @@
 import os
-from typing import Optional
+from typing import Any, Optional
 
 from langchain.callbacks.streaming_stdout import StreamingStdOutCallbackHandler
 from langchain.schema import HumanMessage, SystemMessage
@@ -22,9 +22,27 @@ class GroqLlm(BaseLlm):
         if not self.config.api_key and "GROQ_API_KEY" not in os.environ:
             raise ValueError("Please set the GROQ_API_KEY environment variable or pass it in the config.")
 
-    def get_llm_model_answer(self, prompt) -> str:
-        response = self._get_answer(prompt, self.config)
-        return response
+    def get_llm_model_answer(self, prompt) -> tuple[str, Optional[dict[str, Any]]]:
+        if self.config.token_usage:
+            response, token_info = self._get_answer(prompt, self.config)
+            model_name = "groq/" + self.config.model
+            if model_name not in self.config.model_pricing_map:
+                raise ValueError(
+                    f"Model {model_name} not found in `model_prices_and_context_window.json`. \
+                    You can disable token usage by setting `token_usage` to False."
+                )
+            total_cost = (
+                self.config.model_pricing_map[model_name]["input_cost_per_token"] * token_info["prompt_tokens"]
+            ) + self.config.model_pricing_map[model_name]["output_cost_per_token"] * token_info["completion_tokens"]
+            response_token_info = {
+                "prompt_tokens": token_info["prompt_tokens"],
+                "completion_tokens": token_info["completion_tokens"],
+                "total_tokens": token_info["prompt_tokens"] + token_info["completion_tokens"],
+                "total_cost": round(total_cost, 10),
+                "cost_currency": "USD",
+            }
+            return response, response_token_info
+        return self._get_answer(prompt, self.config)
 
     def _get_answer(self, prompt: str, config: BaseLlmConfig) -> str:
         messages = []
@@ -42,4 +60,8 @@ class GroqLlm(BaseLlm):
             chat = ChatGroq(**kwargs, streaming=config.stream, callbacks=callbacks, api_key=api_key)
         else:
             chat = ChatGroq(**kwargs)
-        return chat.invoke(messages).content
+
+        chat_response = chat.invoke(prompt)
+        if self.config.token_usage:
+            return chat_response.content, chat_response.response_metadata["token_usage"]
+        return chat_response.content

embedchain/llm/mistralai.py

@@ -1,5 +1,5 @@
 import os
-from typing import Optional
+from typing import Any, Optional
 
 from embedchain.config import BaseLlmConfig
 from embedchain.helpers.json_serializable import register_deserializable
@@ -13,8 +13,27 @@ class MistralAILlm(BaseLlm):
         if not self.config.api_key and "MISTRAL_API_KEY" not in os.environ:
             raise ValueError("Please set the MISTRAL_API_KEY environment variable or pass it in the config.")
 
-    def get_llm_model_answer(self, prompt):
-        return MistralAILlm._get_answer(prompt=prompt, config=self.config)
+    def get_llm_model_answer(self, prompt) -> tuple[str, Optional[dict[str, Any]]]:
+        if self.config.token_usage:
+            response, token_info = self._get_answer(prompt, self.config)
+            model_name = "mistralai/" + self.config.model
+            if model_name not in self.config.model_pricing_map:
+                raise ValueError(
+                    f"Model {model_name} not found in `model_prices_and_context_window.json`. \
+                    You can disable token usage by setting `token_usage` to False."
+                )
+            total_cost = (
+                self.config.model_pricing_map[model_name]["input_cost_per_token"] * token_info["prompt_tokens"]
+            ) + self.config.model_pricing_map[model_name]["output_cost_per_token"] * token_info["completion_tokens"]
+            response_token_info = {
+                "prompt_tokens": token_info["prompt_tokens"],
+                "completion_tokens": token_info["completion_tokens"],
+                "total_tokens": token_info["prompt_tokens"] + token_info["completion_tokens"],
+                "total_cost": round(total_cost, 10),
+                "cost_currency": "USD",
+            }
+            return response, response_token_info
+        return self._get_answer(prompt, self.config)
 
     @staticmethod
     def _get_answer(prompt: str, config: BaseLlmConfig):
@@ -47,6 +66,7 @@ class MistralAILlm(BaseLlm):
                 answer += chunk.content
             return answer
         else:
-            response = client.invoke(**kwargs, input=messages)
-            answer = response.content
-            return answer
+            chat_response = client.invoke(**kwargs, input=messages)
+            if config.token_usage:
+                return chat_response.content, chat_response.response_metadata["token_usage"]
+            return chat_response.content

embedchain/llm/nvidia.py

@@ -1,6 +1,6 @@
 import os
 from collections.abc import Iterable
-from typing import Optional, Union
+from typing import Any, Optional, Union
 
 from langchain.callbacks.manager import CallbackManager
 from langchain.callbacks.stdout import StdOutCallbackHandler
@@ -25,8 +25,27 @@ class NvidiaLlm(BaseLlm):
         if not self.config.api_key and "NVIDIA_API_KEY" not in os.environ:
             raise ValueError("Please set the NVIDIA_API_KEY environment variable or pass it in the config.")
 
-    def get_llm_model_answer(self, prompt):
-        return self._get_answer(prompt=prompt, config=self.config)
+    def get_llm_model_answer(self, prompt) -> tuple[str, Optional[dict[str, Any]]]:
+        if self.config.token_usage:
+            response, token_info = self._get_answer(prompt, self.config)
+            model_name = "nvidia/" + self.config.model
+            if model_name not in self.config.model_pricing_map:
+                raise ValueError(
+                    f"Model {model_name} not found in `model_prices_and_context_window.json`. \
+                    You can disable token usage by setting `token_usage` to False."
+                )
+            total_cost = (
+                self.config.model_pricing_map[model_name]["input_cost_per_token"] * token_info["input_tokens"]
+            ) + self.config.model_pricing_map[model_name]["output_cost_per_token"] * token_info["output_tokens"]
+            response_token_info = {
+                "prompt_tokens": token_info["input_tokens"],
+                "completion_tokens": token_info["output_tokens"],
+                "total_tokens": token_info["input_tokens"] + token_info["output_tokens"],
+                "total_cost": round(total_cost, 10),
+                "cost_currency": "USD",
+            }
+            return response, response_token_info
+        return self._get_answer(prompt, self.config)
 
     @staticmethod
     def _get_answer(prompt: str, config: BaseLlmConfig) -> Union[str, Iterable]:
@@ -43,4 +62,7 @@ class NvidiaLlm(BaseLlm):
         if labels:
             params["labels"] = labels
         llm = ChatNVIDIA(**params, callback_manager=CallbackManager(callback_manager))
-        return llm.invoke(prompt).content if labels is None else llm.invoke(prompt, labels=labels).content
+        chat_response = llm.invoke(prompt) if labels is None else llm.invoke(prompt, labels=labels)
+        if config.token_usage:
+            return chat_response.content, chat_response.response_metadata["token_usage"]
+        return chat_response.content

embedchain/llm/openai.py

@@ -23,9 +23,28 @@ class OpenAILlm(BaseLlm):
         self.tools = tools
         super().__init__(config=config)
 
-    def get_llm_model_answer(self, prompt) -> str:
-        response = self._get_answer(prompt, self.config)
-        return response
+    def get_llm_model_answer(self, prompt) -> tuple[str, Optional[dict[str, Any]]]:
+        if self.config.token_usage:
+            response, token_info = self._get_answer(prompt, self.config)
+            model_name = "openai/" + self.config.model
+            if model_name not in self.config.model_pricing_map:
+                raise ValueError(
+                    f"Model {model_name} not found in `model_prices_and_context_window.json`. \
+                    You can disable token usage by setting `token_usage` to False."
+                )
+            total_cost = (
+                self.config.model_pricing_map[model_name]["input_cost_per_token"] * token_info["prompt_tokens"]
+            ) + self.config.model_pricing_map[model_name]["output_cost_per_token"] * token_info["completion_tokens"]
+            response_token_info = {
+                "prompt_tokens": token_info["prompt_tokens"],
+                "completion_tokens": token_info["completion_tokens"],
+                "total_tokens": token_info["prompt_tokens"] + token_info["completion_tokens"],
+                "total_cost": round(total_cost, 10),
+                "cost_currency": "USD",
+            }
+            return response, response_token_info
+
+        return self._get_answer(prompt, self.config)
 
     def _get_answer(self, prompt: str, config: BaseLlmConfig) -> str:
         messages = []
@@ -66,7 +85,10 @@ class OpenAILlm(BaseLlm):
         if self.tools:
             return self._query_function_call(chat, self.tools, messages)
 
-        return chat.invoke(messages).content
+        chat_response = chat.invoke(messages)
+        if self.config.token_usage:
+            return chat_response.content, chat_response.response_metadata["token_usage"]
+        return chat_response.content
 
     def _query_function_call(
         self,

embedchain/llm/together.py

@@ -1,8 +1,13 @@
 import importlib
 import os
-from typing import Optional
+from typing import Any, Optional
 
-from langchain_community.llms import Together
+try:
+    from langchain_together import ChatTogether
+except ImportError:
+    raise ImportError(
+        "Please install the langchain_together package by running `pip install langchain_together==0.1.3`."
+    )
 
 from embedchain.config import BaseLlmConfig
 from embedchain.helpers.json_serializable import register_deserializable
@@ -24,20 +29,43 @@ class TogetherLlm(BaseLlm):
         if not self.config.api_key and "TOGETHER_API_KEY" not in os.environ:
             raise ValueError("Please set the TOGETHER_API_KEY environment variable or pass it in the config.")
 
-    def get_llm_model_answer(self, prompt):
+    def get_llm_model_answer(self, prompt) -> tuple[str, Optional[dict[str, Any]]]:
         if self.config.system_prompt:
             raise ValueError("TogetherLlm does not support `system_prompt`")
-        return TogetherLlm._get_answer(prompt=prompt, config=self.config)
+
+        if self.config.token_usage:
+            response, token_info = self._get_answer(prompt, self.config)
+            model_name = "together/" + self.config.model
+            if model_name not in self.config.model_pricing_map:
+                raise ValueError(
+                    f"Model {model_name} not found in `model_prices_and_context_window.json`. \
+                    You can disable token usage by setting `token_usage` to False."
+                )
+            total_cost = (
+                self.config.model_pricing_map[model_name]["input_cost_per_token"] * token_info["prompt_tokens"]
+            ) + self.config.model_pricing_map[model_name]["output_cost_per_token"] * token_info["completion_tokens"]
+            response_token_info = {
+                "prompt_tokens": token_info["prompt_tokens"],
+                "completion_tokens": token_info["completion_tokens"],
+                "total_tokens": token_info["prompt_tokens"] + token_info["completion_tokens"],
+                "total_cost": round(total_cost, 10),
+                "cost_currency": "USD",
+            }
+            return response, response_token_info
+        return self._get_answer(prompt, self.config)
 
     @staticmethod
     def _get_answer(prompt: str, config: BaseLlmConfig) -> str:
-        api_key = config.api_key or os.getenv("TOGETHER_API_KEY")
-        llm = Together(
-            together_api_key=api_key,
-            model=config.model,
-            max_tokens=config.max_tokens,
-            temperature=config.temperature,
-            top_p=config.top_p,
-        )
+        api_key = config.api_key or os.environ["TOGETHER_API_KEY"]
+        kwargs = {
+            "model_name": config.model or "mixtral-8x7b-32768",
+            "temperature": config.temperature,
+            "max_tokens": config.max_tokens,
+            "together_api_key": api_key,
+        }
 
-        return llm.invoke(prompt)
+        chat = ChatTogether(**kwargs)
+        chat_response = chat.invoke(prompt)
+        if config.token_usage:
+            return chat_response.content, chat_response.response_metadata["token_usage"]
+        return chat_response.content

embedchain/llm/vertex_ai.py

@@ -1,6 +1,6 @@
 import importlib
 import logging
-from typing import Optional
+from typing import Any, Optional
 
 from langchain.callbacks.streaming_stdout import StreamingStdOutCallbackHandler
 from langchain_google_vertexai import ChatVertexAI
@@ -24,16 +24,35 @@ class VertexAILlm(BaseLlm):
             ) from None
         super().__init__(config=config)
 
-    def get_llm_model_answer(self, prompt):
-        return VertexAILlm._get_answer(prompt=prompt, config=self.config)
+    def get_llm_model_answer(self, prompt) -> tuple[str, Optional[dict[str, Any]]]:
+        if self.config.token_usage:
+            response, token_info = self._get_answer(prompt, self.config)
+            model_name = "vertexai/" + self.config.model
+            if model_name not in self.config.model_pricing_map:
+                raise ValueError(
+                    f"Model {model_name} not found in `model_prices_and_context_window.json`. \
+                    You can disable token usage by setting `token_usage` to False."
+                )
+            total_cost = (
+                self.config.model_pricing_map[model_name]["input_cost_per_token"] * token_info["prompt_token_count"]
+            ) + self.config.model_pricing_map[model_name]["output_cost_per_token"] * token_info[
+                "candidates_token_count"
+            ]
+            response_token_info = {
+                "prompt_tokens": token_info["prompt_token_count"],
+                "completion_tokens": token_info["candidates_token_count"],
+                "total_tokens": token_info["prompt_token_count"] + token_info["candidates_token_count"],
+                "total_cost": round(total_cost, 10),
+                "cost_currency": "USD",
+            }
+            return response, response_token_info
+        return self._get_answer(prompt, self.config)
 
     @staticmethod
     def _get_answer(prompt: str, config: BaseLlmConfig) -> str:
         if config.top_p and config.top_p != 1:
             logger.warning("Config option `top_p` is not supported by this model.")
 
-        messages = BaseLlm._get_messages(prompt, system_prompt=config.system_prompt)
-
         if config.stream:
             callbacks = config.callbacks if config.callbacks else [StreamingStdOutCallbackHandler()]
             llm = ChatVertexAI(
@@ -42,4 +61,8 @@ class VertexAILlm(BaseLlm):
         else:
             llm = ChatVertexAI(temperature=config.temperature, model=config.model)
 
-        return llm.invoke(messages).content
+        messages = VertexAILlm._get_messages(prompt)
+        chat_response = llm.invoke(messages)
+        if config.token_usage:
+            return chat_response.content, chat_response.response_metadata["usage_metadata"]
+        return chat_response.content

embedchain/utils/misc.py

@@ -428,6 +428,7 @@ def validate_config(config_data):
                     Optional("top_p"): Or(float, int),
                     Optional("stream"): bool,
                     Optional("online"): bool,
+                    Optional("token_usage"): bool,
                     Optional("template"): str,
                     Optional("prompt"): str,
                     Optional("system_prompt"): str,