langroid 0.51.1__tar.gz → 0.52.0__tar.gz

{langroid-0.51.1 → langroid-0.52.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: langroid
-Version: 0.51.1
+Version: 0.52.0
 Summary: Harness LLMs with Multi-Agent Programming
 Author-email: Prasad Chalasani <pchalasani@gmail.com>
 License: MIT
@@ -846,7 +846,7 @@ import langroid.language_models as lm
 
 mdl = lm.OpenAIGPT(
     lm.OpenAIGPTConfig(
-        chat_model=lm.OpenAIChatModel.GPT4, # or, e.g.  "ollama/qwen2.5"
+        chat_model=lm.OpenAIChatModel.GPT4o, # or, e.g.  "ollama/qwen2.5"
     ),
 )
 

{langroid-0.51.1 → langroid-0.52.0}/README.md

@@ -644,7 +644,7 @@ import langroid.language_models as lm
 
 mdl = lm.OpenAIGPT(
     lm.OpenAIGPTConfig(
-        chat_model=lm.OpenAIChatModel.GPT4, # or, e.g.  "ollama/qwen2.5"
+        chat_model=lm.OpenAIChatModel.GPT4o, # or, e.g.  "ollama/qwen2.5"
     ),
 )
 

{langroid-0.51.1 → langroid-0.52.0}/langroid/agent/base.py

@@ -47,6 +47,7 @@ from langroid.language_models.base import (
 )
 from langroid.language_models.openai_gpt import OpenAIGPT, OpenAIGPTConfig
 from langroid.mytypes import Entity
+from langroid.parsing.file_attachment import FileAttachment
 from langroid.parsing.parse_json import extract_top_level_json
 from langroid.parsing.parser import Parser, ParsingConfig
 from langroid.prompts.prompts_config import PromptsConfig
@@ -440,6 +441,7 @@ class Agent(ABC):
     def create_agent_response(
         self,
         content: str | None = None,
+        files: List[FileAttachment] = [],
         content_any: Any = None,
         tool_messages: List[ToolMessage] = [],
         oai_tool_calls: Optional[List[OpenAIToolCall]] = None,
@@ -452,6 +454,7 @@ class Agent(ABC):
         return self.response_template(
             Entity.AGENT,
             content=content,
+            files=files,
             content_any=content_any,
             tool_messages=tool_messages,
             oai_tool_calls=oai_tool_calls,
@@ -689,6 +692,7 @@ class Agent(ABC):
         self,
         e: Entity,
         content: str | None = None,
+        files: List[FileAttachment] = [],
         content_any: Any = None,
         tool_messages: List[ToolMessage] = [],
         oai_tool_calls: Optional[List[OpenAIToolCall]] = None,
@@ -700,6 +704,7 @@ class Agent(ABC):
         """Template for response from entity `e`."""
         return ChatDocument(
             content=content or "",
+            files=files,
             content_any=content_any,
             tool_messages=tool_messages,
             oai_tool_calls=oai_tool_calls,
@@ -714,6 +719,7 @@ class Agent(ABC):
     def create_user_response(
         self,
         content: str | None = None,
+        files: List[FileAttachment] = [],
         content_any: Any = None,
         tool_messages: List[ToolMessage] = [],
         oai_tool_calls: List[OpenAIToolCall] | None = None,
@@ -726,6 +732,7 @@ class Agent(ABC):
         return self.response_template(
             e=Entity.USER,
             content=content,
+            files=files,
             content_any=content_any,
             tool_messages=tool_messages,
             oai_tool_calls=oai_tool_calls,

{langroid-0.51.1 → langroid-0.52.0}/langroid/agent/chat_agent.py

@@ -1511,12 +1511,14 @@ class ChatAgent(Agent):
         output_len = self.config.llm.model_max_output_tokens
         if (
             truncate
-            and self.chat_num_tokens(hist)
-            > self.llm.chat_context_length() - self.config.llm.model_max_output_tokens
+            and output_len > self.llm.chat_context_length() - self.chat_num_tokens(hist)
         ):
             # chat + output > max context length,
-            # so first try to shorten requested output len to fit.
-            output_len = self.llm.chat_context_length() - self.chat_num_tokens(hist)
+            # so first try to shorten requested output len to fit;
+            # use an extra margin of 300 tokens in case our calcs are off
+            output_len = (
+                self.llm.chat_context_length() - self.chat_num_tokens(hist) - 300
+            )
             if output_len < self.config.llm.min_output_tokens:
                 # unacceptably small output len, so drop early parts of conv history
                 # if output_len is still too long, then drop early parts of conv history
@@ -1534,10 +1536,17 @@ class ChatAgent(Agent):
                         # and last message (user msg).
                         raise ValueError(
                             """
-                        The message history is longer than the max chat context 
-                        length allowed, and we have run out of messages to drop.
-                        HINT: In your `OpenAIGPTConfig` object, try increasing
-                        `chat_context_length` or decreasing `model_max_output_tokens`.
+                        The (message history + max_output_tokens) is longer than the 
+                        max chat context length of this model, and we have tried 
+                        reducing the requested max output tokens, as well as dropping 
+                        early parts of the message history, to accommodate the model 
+                        context length, but we have run out of msgs to drop.
+                         
+                        HINT: In the `llm` field of your `ChatAgentConfig` object, 
+                        which is of type `LLMConfig/OpenAIGPTConfig`, try 
+                        - increasing `chat_context_length` 
+                            (if accurate for the model), or  
+                        - decreasing `max_output_tokens`
                         """
                         )
                     # drop the second message, i.e. first msg after the sys msg

{langroid-0.51.1 → langroid-0.52.0}/langroid/agent/chat_document.py

@@ -19,6 +19,7 @@ from langroid.language_models.base import (
 )
 from langroid.mytypes import DocMetaData, Document, Entity
 from langroid.parsing.agent_chats import parse_message
+from langroid.parsing.file_attachment import FileAttachment
 from langroid.parsing.parse_json import extract_top_level_json, top_level_json_field
 from langroid.pydantic_v1 import BaseModel, Extra
 from langroid.utils.object_registry import ObjectRegistry
@@ -119,6 +120,7 @@ class ChatDocument(Document):
 
     reasoning: str = ""  # reasoning produced by a reasoning LLM
     content_any: Any = None  # to hold arbitrary data returned by responders
+    files: List[FileAttachment] = []  # list of file attachments
     oai_tool_calls: Optional[List[OpenAIToolCall]] = None
     oai_tool_id2result: Optional[OrderedDict[str, str]] = None
     oai_tool_choice: ToolChoiceTypes | Dict[str, Dict[str, str] | str] = "auto"
@@ -356,12 +358,8 @@ class ChatDocument(Document):
         Returns:
             List[LLMMessage]: list of LLMMessages corresponding to this ChatDocument.
         """
-        sender_name = None
+
         sender_role = Role.USER
-        fun_call = None
-        oai_tool_calls = None
-        tool_id = ""  # for OpenAI Assistant
-        chat_document_id: str = ""
         if isinstance(message, str):
             message = ChatDocument.from_str(message)
         content = message.content or to_string(message.content_any) or ""
@@ -381,6 +379,8 @@ class ChatDocument(Document):
             # same reasoning as for function-call above
             content += " " + "\n\n".join(str(tc) for tc in oai_tool_calls)
             oai_tool_calls = None
+        # some LLM APIs (e.g. gemini) don't like empty msg
+        content = content or " "
         sender_name = message.metadata.sender_name
         tool_ids = message.metadata.tool_ids
         tool_id = tool_ids[-1] if len(tool_ids) > 0 else ""
@@ -409,6 +409,7 @@ class ChatDocument(Document):
                         role=Role.TOOL,
                         tool_call_id=oai_tools[0].id,
                         content=content,
+                        files=message.files,
                         chat_document_id=chat_document_id,
                     )
                 ]
@@ -424,6 +425,7 @@ class ChatDocument(Document):
                         role=Role.TOOL,
                         tool_call_id=message.metadata.oai_tool_id,
                         content=content,
+                        files=message.files,
                         chat_document_id=chat_document_id,
                     )
                 ]
@@ -437,7 +439,8 @@ class ChatDocument(Document):
                     LLMMessage(
                         role=Role.TOOL,
                         tool_call_id=tool_id,
-                        content=result,
+                        content=result or " ",
+                        files=message.files,
                         chat_document_id=chat_document_id,
                     )
                     for tool_id, result in message.oai_tool_id2result.items()
@@ -450,6 +453,7 @@ class ChatDocument(Document):
                 role=sender_role,
                 tool_id=tool_id,  # for OpenAI Assistant
                 content=content,
+                files=message.files,
                 function_call=fun_call,
                 tool_calls=oai_tool_calls,
                 name=sender_name,

{langroid-0.51.1 → langroid-0.52.0}/langroid/agent/special/doc_chat_agent.py

@@ -204,8 +204,8 @@ class DocChatAgentConfig(ChatAgentConfig):
 
     llm: OpenAIGPTConfig = OpenAIGPTConfig(
         type="openai",
-        chat_model=OpenAIChatModel.GPT4,
-        completion_model=OpenAIChatModel.GPT4,
+        chat_model=OpenAIChatModel.GPT4o,
+        completion_model=OpenAIChatModel.GPT4o,
         timeout=40,
     )
     prompts: PromptsConfig = PromptsConfig(

{langroid-0.51.1 → langroid-0.52.0}/langroid/agent/special/table_chat_agent.py

@@ -118,8 +118,8 @@ class TableChatAgentConfig(ChatAgentConfig):
     vecdb: None | VectorStoreConfig = None
     llm: OpenAIGPTConfig = OpenAIGPTConfig(
         type="openai",
-        chat_model=OpenAIChatModel.GPT4,
-        completion_model=OpenAIChatModel.GPT4,
+        chat_model=OpenAIChatModel.GPT4o,
+        completion_model=OpenAIChatModel.GPT4o,
     )
     prompts: PromptsConfig = PromptsConfig(
         max_tokens=1000,

{langroid-0.51.1 → langroid-0.52.0}/langroid/language_models/base.py

@@ -21,6 +21,7 @@ from langroid.cachedb.base import CacheDBConfig
 from langroid.cachedb.redis_cachedb import RedisCacheConfig
 from langroid.language_models.model_info import ModelInfo, get_model_info
 from langroid.parsing.agent_chats import parse_message
+from langroid.parsing.file_attachment import FileAttachment
 from langroid.parsing.parse_json import parse_imperfect_json, top_level_json_field
 from langroid.prompts.dialog import collate_chat_history
 from langroid.pydantic_v1 import BaseModel, BaseSettings, Field
@@ -53,6 +54,13 @@ class StreamEventType(Enum):
     TOOL_ARGS = 5
 
 
+class RetryParams(BaseSettings):
+    max_retries: int = 5
+    initial_delay: float = 1.0
+    exponential_base: float = 1.3
+    jitter: bool = True
+
+
 class LLMConfig(BaseSettings):
     """
     Common configuration for all language models.
@@ -63,7 +71,8 @@ class LLMConfig(BaseSettings):
     streamer_async: Optional[Callable[..., Awaitable[None]]] = async_noop_fn
     api_base: str | None = None
     formatter: None | str = None
-    max_output_tokens: int | None = 8192  # specify None to use model_max_output_tokens
+    # specify None if you want to use the full max output tokens of the model
+    max_output_tokens: int | None = 8192
     timeout: int = 20  # timeout for API requests
     chat_model: str = ""
     completion_model: str = ""
@@ -86,11 +95,13 @@ class LLMConfig(BaseSettings):
     # Dict of model -> (input/prompt cost, output/completion cost)
     chat_cost_per_1k_tokens: Tuple[float, float] = (0.0, 0.0)
     completion_cost_per_1k_tokens: Tuple[float, float] = (0.0, 0.0)
+    retry_params: RetryParams = RetryParams()
 
     @property
     def model_max_output_tokens(self) -> int:
-        return (
-            self.max_output_tokens or get_model_info(self.chat_model).max_output_tokens
+        return min(
+            self.max_output_tokens or get_model_info(self.chat_model).max_output_tokens,
+            get_model_info(self.chat_model).max_output_tokens,
         )
 
 
@@ -263,13 +274,14 @@ class LLMMessage(BaseModel):
     tool_call_id: Optional[str] = None  # which OpenAI LLM tool this is a response to
     tool_id: str = ""  # used by OpenAIAssistant
     content: str
+    files: List[FileAttachment] = []
     function_call: Optional[LLMFunctionCall] = None
     tool_calls: Optional[List[OpenAIToolCall]] = None
     timestamp: datetime = Field(default_factory=datetime.utcnow)
     # link to corresponding chat document, for provenance/rewind purposes
     chat_document_id: str = ""
 
-    def api_dict(self, has_system_role: bool = True) -> Dict[str, Any]:
+    def api_dict(self, model: str, has_system_role: bool = True) -> Dict[str, Any]:
         """
         Convert to dictionary for API request, keeping ONLY
         the fields that are expected in an API call!
@@ -283,6 +295,17 @@ class LLMMessage(BaseModel):
             dict: dictionary representation of LLM message
         """
         d = self.dict()
+        files: List[FileAttachment] = d.pop("files")
+        if len(files) > 0 and self.role == Role.USER:
+            # In there are files, then content is an array of
+            # different content-parts
+            d["content"] = [
+                dict(
+                    type="text",
+                    text=self.content,
+                )
+            ] + [f.to_dict(model) for f in self.files]
+
         # if there is a key k = "role" with value "system", change to "user"
         # in case has_system_role is False
         if not has_system_role and "role" in d and d["role"] == "system":

{langroid-0.51.1 → langroid-0.52.0}/langroid/language_models/model_info.py

@@ -24,13 +24,16 @@ class OpenAIChatModel(ModelName):
     """Enum for OpenAI Chat models"""
 
     GPT3_5_TURBO = "gpt-3.5-turbo-1106"
-    GPT4 = "gpt-4"
+    GPT4 = "gpt-4o"  # avoid deprecated gpt-4
     GPT4_TURBO = "gpt-4-turbo"
     GPT4o = "gpt-4o"
     GPT4o_MINI = "gpt-4o-mini"
     O1 = "o1"
     O1_MINI = "o1-mini"
     O3_MINI = "o3-mini"
+    GPT4_1 = "gpt-4.1"
+    GPT4_1_MINI = "gpt-4.1-mini"
+    GPT4_1_NANO = "gpt-4.1-nano"
 
 
 class OpenAICompletionModel(str, Enum):
@@ -44,6 +47,7 @@ class AnthropicModel(ModelName):
     """Enum for Anthropic models"""
 
     CLAUDE_3_5_SONNET = "claude-3-5-sonnet-latest"
+    CLAUDE_3_7_SONNET = "claude-3-7-sonnet-latest"
     CLAUDE_3_OPUS = "claude-3-opus-latest"
     CLAUDE_3_SONNET = "claude-3-sonnet-20240229"
     CLAUDE_3_HAIKU = "claude-3-haiku-20240307"
@@ -63,6 +67,7 @@ class GeminiModel(ModelName):
     GEMINI_1_5_FLASH = "gemini-1.5-flash"
     GEMINI_1_5_FLASH_8B = "gemini-1.5-flash-8b"
     GEMINI_1_5_PRO = "gemini-1.5-pro"
+    GEMINI_2_5_PRO = "gemini-2.5-pro-exp-02-05"
     GEMINI_2_PRO = "gemini-2.0-pro-exp-02-05"
     GEMINI_2_FLASH = "gemini-2.0-flash"
     GEMINI_2_FLASH_LITE = "gemini-2.0-flash-lite-preview"
@@ -160,6 +165,33 @@ MODEL_INFO: Dict[str, ModelInfo] = {
         output_cost_per_million=30.0,
         description="GPT-4 Turbo",
     ),
+    OpenAIChatModel.GPT4_1_NANO.value: ModelInfo(
+        name=OpenAIChatModel.GPT4_1_NANO.value,
+        provider=ModelProvider.OPENAI,
+        context_length=1_047_576,
+        max_output_tokens=32_768,
+        input_cost_per_million=0.10,
+        output_cost_per_million=0.40,
+        description="GPT-4.1",
+    ),
+    OpenAIChatModel.GPT4_1_MINI.value: ModelInfo(
+        name=OpenAIChatModel.GPT4_1_MINI.value,
+        provider=ModelProvider.OPENAI,
+        context_length=1_047_576,
+        max_output_tokens=32_768,
+        input_cost_per_million=0.40,
+        output_cost_per_million=1.60,
+        description="GPT-4.1 Mini",
+    ),
+    OpenAIChatModel.GPT4_1.value: ModelInfo(
+        name=OpenAIChatModel.GPT4_1.value,
+        provider=ModelProvider.OPENAI,
+        context_length=1_047_576,
+        max_output_tokens=32_768,
+        input_cost_per_million=2.00,
+        output_cost_per_million=8.00,
+        description="GPT-4.1",
+    ),
     OpenAIChatModel.GPT4o.value: ModelInfo(
         name=OpenAIChatModel.GPT4o.value,
         provider=ModelProvider.OPENAI,

{langroid-0.51.1 → langroid-0.52.0}/langroid/language_models/openai_gpt.py

@@ -91,10 +91,13 @@ LLAMACPP_API_KEY = os.environ.get("LLAMA_API_KEY", DUMMY_API_KEY)
 
 openai_chat_model_pref_list = [
     OpenAIChatModel.GPT4o,
+    OpenAIChatModel.GPT4_1_NANO,
+    OpenAIChatModel.GPT4_1_MINI,
+    OpenAIChatModel.GPT4_1,
     OpenAIChatModel.GPT4o_MINI,
     OpenAIChatModel.O1_MINI,
+    OpenAIChatModel.O3_MINI,
     OpenAIChatModel.O1,
-    OpenAIChatModel.GPT3_5_TURBO,
 ]
 
 openai_completion_model_pref_list = [
@@ -1731,8 +1734,7 @@ class OpenAIGPT(LanguageModel):
             logging.error(friendly_error(e, "Error in OpenAIGPT.achat: "))
             raise e
 
-    @retry_with_exponential_backoff
-    def _chat_completions_with_backoff(self, **kwargs):  # type: ignore
+    def _chat_completions_with_backoff_body(self, **kwargs):  # type: ignore
         cached = False
         hashed_key, result = self._cache_lookup("Completion", **kwargs)
         if result is not None:
@@ -1781,8 +1783,17 @@ class OpenAIGPT(LanguageModel):
                 self._cache_store(hashed_key, result.model_dump())
         return cached, hashed_key, result
 
-    @async_retry_with_exponential_backoff
-    async def _achat_completions_with_backoff(self, **kwargs):  # type: ignore
+    def _chat_completions_with_backoff(self, **kwargs):  # type: ignore
+        retry_func = retry_with_exponential_backoff(
+            self._chat_completions_with_backoff_body,
+            initial_delay=self.config.retry_params.initial_delay,
+            max_retries=self.config.retry_params.max_retries,
+            exponential_base=self.config.retry_params.exponential_base,
+            jitter=self.config.retry_params.jitter,
+        )
+        return retry_func(**kwargs)
+
+    async def _achat_completions_with_backoff_body(self, **kwargs):  # type: ignore
         cached = False
         hashed_key, result = self._cache_lookup("Completion", **kwargs)
         if result is not None:
@@ -1836,6 +1847,16 @@ class OpenAIGPT(LanguageModel):
                 self._cache_store(hashed_key, result.model_dump())
         return cached, hashed_key, result
 
+    async def _achat_completions_with_backoff(self, **kwargs):  # type: ignore
+        retry_func = async_retry_with_exponential_backoff(
+            self._achat_completions_with_backoff_body,
+            initial_delay=self.config.retry_params.initial_delay,
+            max_retries=self.config.retry_params.max_retries,
+            exponential_base=self.config.retry_params.exponential_base,
+            jitter=self.config.retry_params.jitter,
+        )
+        return await retry_func(**kwargs)
+
     def _prep_chat_completion(
         self,
         messages: Union[str, List[LLMMessage]],
@@ -1876,10 +1897,13 @@ class OpenAIGPT(LanguageModel):
         args: Dict[str, Any] = dict(
             model=chat_model,
             messages=[
-                m.api_dict(has_system_role=self.info().allows_system_message)
+                m.api_dict(
+                    self.config.chat_model,
+                    has_system_role=self.info().allows_system_message,
+                )
                 for m in (llm_messages)
             ],
-            max_tokens=max_tokens,
+            max_completion_tokens=max_tokens,
             stream=self.get_stream(),
         )
         if self.get_stream():
@@ -2073,7 +2097,7 @@ class OpenAIGPT(LanguageModel):
             function_call,
             response_format,
         )
-        cached, hashed_key, response = self._chat_completions_with_backoff(**args)
+        cached, hashed_key, response = self._chat_completions_with_backoff(**args)  # type: ignore
         if self.get_stream() and not cached:
             llm_response, openai_response = self._stream_response(response, chat=True)
             self._cache_store(hashed_key, openai_response)
@@ -2106,7 +2130,7 @@ class OpenAIGPT(LanguageModel):
             function_call,
             response_format,
         )
-        cached, hashed_key, response = await self._achat_completions_with_backoff(
+        cached, hashed_key, response = await self._achat_completions_with_backoff(  # type: ignore
             **args
         )
         if self.get_stream() and not cached:

{langroid-0.51.1 → langroid-0.52.0}/langroid/parsing/document_parser.py

@@ -31,7 +31,7 @@ if TYPE_CHECKING:
     from PIL import Image
 
 from langroid.mytypes import DocMetaData, Document
-from langroid.parsing.parser import Parser, ParsingConfig
+from langroid.parsing.parser import LLMPdfParserConfig, Parser, ParsingConfig
 
 logger = logging.getLogger(__name__)
 
@@ -1040,7 +1040,8 @@ class LLMPdfParser(DocumentParser):
             raise ValueError(
                 "LLMPdfParser requires a llm-based config in pdf parsing config"
             )
-        self.model_name = config.pdf.llm_parser_config.model_name
+        self.llm_parser_config: LLMPdfParserConfig = config.pdf.llm_parser_config
+        self.model_name = self.llm_parser_config.model_name
 
         # Ensure output directory exists
         self.OUTPUT_DIR.mkdir(parents=True, exist_ok=True)
@@ -1059,9 +1060,7 @@ class LLMPdfParser(DocumentParser):
         temp_file.close()
         self.output_filename = Path(temp_file.name)
 
-        self.max_tokens = (
-            config.pdf.llm_parser_config.max_tokens or self.DEFAULT_MAX_TOKENS
-        )
+        self.max_tokens = self.llm_parser_config.max_tokens or self.DEFAULT_MAX_TOKENS
 
         """
         If True, each PDF page is processed as a separate chunk,
@@ -1069,12 +1068,12 @@ class LLMPdfParser(DocumentParser):
         grouped into chunks based on `max_token_limit` before being sent
         to the LLM.
         """
-        self.split_on_page = config.pdf.llm_parser_config.split_on_page or False
+        self.split_on_page = self.llm_parser_config.split_on_page or False
 
         # Rate limiting parameters
         import asyncio
 
-        self.requests_per_minute = config.pdf.llm_parser_config.requests_per_minute or 5
+        self.requests_per_minute = self.llm_parser_config.requests_per_minute or 5
 
         """
         A semaphore to control the number of concurrent requests to the LLM,
@@ -1231,6 +1230,7 @@ class LLMPdfParser(DocumentParser):
                     llm_config = OpenAIGPTConfig(
                         chat_model=self.model_name,
                         max_output_tokens=self.max_tokens,
+                        timeout=self.llm_parser_config.timeout,
                     )
                     llm = OpenAIGPT(config=llm_config)
                     page_nums = self._page_num_str(chunk.get("page_numbers", "?"))
@@ -1242,7 +1242,7 @@ class LLMPdfParser(DocumentParser):
                             image_url=dict(url=data_uri),
                         )
                     elif "claude" in self.model_name.lower():
-                        # optimistrally try this: some API proxies like litellm
+                        # optimistically try this: some API proxies like litellm
                         # support this, and others may not.
                         file_content = dict(
                             type="file",
@@ -1259,27 +1259,32 @@ class LLMPdfParser(DocumentParser):
                                 file_data=data_uri,
                             ),
                         )
+                    prompt = (
+                        self.llm_parser_config.prompt
+                        or self.LLM_PDF_MD_SYSTEM_INSTRUCTION
+                    )
+                    system_prompt = (
+                        self.llm_parser_config.system_prompt
+                        or """
+                         You are an expert pdf -> markdown converter.
+                         Do NOT use any triple backquotes when you present the
+                         markdown content,like ```markdown etc.
+                         FAITHFULLY CONVERT THE PDF TO MARKDOWN,
+                         retaining ALL content as you find it.
+                        """
+                    )
 
                     # Send the request with PDF content and system instructions
                     response = await llm.async_client.chat.completions.create(  # type: ignore
                         model=self.model_name.split("/")[-1],
                         messages=[
-                            dict(
-                                role="system",
-                                content="""
-                                You are an expert pdf -> markdown converter.
-                                Do NOT use any triple backquotes when you present the 
-                                markdown content,like ```markdown etc.
-                                FAITHFULLY CONVERT THE PDF TO MARKDOWN, 
-                                retaining ALL content as you find it.
-                                """,
-                            ),
+                            dict(role="system", content=system_prompt),
                             dict(  # type: ignore
                                 role="user",
                                 content=[
                                     dict(
                                         type="text",
-                                        text=self.LLM_PDF_MD_SYSTEM_INSTRUCTION,
+                                        text=prompt,
                                     ),
                                     file_content,
                                 ],