Merge pull request #941 from assafelovic/feature/strategic_llm

Feature/strategic llm

docs/docs/gpt-researcher/gptr/config.md

@@ -22,6 +22,7 @@ Below is a list of current supported options:
 - **`EMBEDDING`**: Embedding model. Defaults to `openai:text-embedding-3-small`. Options: `ollama`, `huggingface`, `azure_openai`, `custom`.
 - **`FAST_LLM`**: Model name for fast LLM operations such summaries. Defaults to `openai:gpt-4o-mini`.
 - **`SMART_LLM`**: Model name for smart operations like generating research reports and reasoning. Defaults to `openai:gpt-4o`.
+- **`STRATEGIC_LLM`**: Model name for strategic operations like generating research plans and strategies. Defaults to `openai:o1-preview`.
 - **`FAST_TOKEN_LIMIT`**: Maximum token limit for fast LLM responses. Defaults to `2000`.
 - **`SMART_TOKEN_LIMIT`**: Maximum token limit for smart LLM responses. Defaults to `4000`.
 - **`BROWSE_CHUNK_MAX_LENGTH`**: Maximum length of text chunks to browse in web sources. Defaults to `8192`.

gpt_researcher/actions/__init__.py

@@ -1,5 +1,5 @@
 from .retriever import get_retriever, get_retrievers
-from .query_processing import get_sub_queries
+from .query_processing import plan_research_outline
 from .agent_creator import extract_json_with_regex, choose_agent
 from .web_scraping import scrape_urls
 from .report_generation import write_conclusion, summarize_url, generate_draft_section_titles, generate_report, write_report_introduction
@@ -9,7 +9,7 @@
 __all__ = [
     "get_retriever",
     "get_retrievers",
-    "get_sub_queries",
+    "plan_research_outline",
     "extract_json_with_regex",
     "scrape_urls",
     "write_conclusion",

gpt_researcher/actions/query_processing.py

@@ -1,58 +1,113 @@
 import json_repair
 from ..utils.llm import create_chat_completion
 from ..prompts import generate_search_queries_prompt
-from typing import Any
+from typing import Any, List, Dict
+from ..config import Config
+import logging
 
+logger = logging.getLogger(__name__)
 
-async def get_sub_queries(
+async def get_search_results(query: str, retriever: Any) -> List[Dict[str, Any]]:
+    """
+    Get web search results for a given query.
+    
+    Args:
+        query: The search query
+        retriever: The retriever instance
+    
+    Returns:
+        A list of search results
+    """
+    search_retriever = retriever(query)
+    return search_retriever.search()
+
+async def generate_sub_queries(
     query: str,
-    retriever: Any,
+    parent_query: str,
+    report_type: str,
+    context: List[Dict[str, Any]],
+    cfg: Config,
+    cost_callback: callable = None
+) -> List[str]:
+    """
+    Generate sub-queries using the specified LLM model.
+    
+    Args:
+        query: The original query
+        parent_query: The parent query
+        report_type: The type of report
+        max_iterations: Maximum number of research iterations
+        context: Search results context
+        cfg: Configuration object
+        cost_callback: Callback for cost calculation
+    
+    Returns:
+        A list of sub-queries
+    """
+    gen_queries_prompt = generate_search_queries_prompt(
+        query,
+        parent_query,
+        report_type,
+        max_iterations=cfg.max_iterations or 1,
+        context=context
+    )
+
+    try:
+        response = await create_chat_completion(
+            model=cfg.strategic_llm_model,
+            messages=[{"role": "user", "content": gen_queries_prompt}],
+            temperature=1,
+            llm_provider=cfg.strategic_llm_provider,
+            max_tokens=None,
+            llm_kwargs=cfg.llm_kwargs,
+            cost_callback=cost_callback,
+        )
+    except Exception as e:
+        logger.warning(f"Error with strategic LLM: {e}. Falling back to smart LLM.")
+        response = await create_chat_completion(
+            model=cfg.smart_llm_model,
+            messages=[{"role": "user", "content": gen_queries_prompt}],
+            temperature=cfg.temperature,
+            max_tokens=cfg.smart_token_limit,
+            llm_provider=cfg.smart_llm_provider,
+            llm_kwargs=cfg.llm_kwargs,
+            cost_callback=cost_callback,
+        )
+
+    return json_repair.loads(response)
+
+async def plan_research_outline(
+    query: str,
+    search_results: List[Dict[str, Any]],
     agent_role_prompt: str,
-    cfg,
+    cfg: Config,
     parent_query: str,
     report_type: str,
     cost_callback: callable = None,
-):
+) -> List[str]:
     """
-    Gets the sub queries
+    Plan the research outline by generating sub-queries.
+    
     Args:
-        query: original query
-        retriever: retriever instance
-        agent_role_prompt: agent role prompt
-        cfg: Config
-        parent_query: parent query
-        report_type: report type
-        cost_callback: callback for cost calculation
-
+        query: Original query
+        retriever: Retriever instance
+        agent_role_prompt: Agent role prompt
+        cfg: Configuration object
+        parent_query: Parent query
+        report_type: Report type
+        cost_callback: Callback for cost calculation
+    
     Returns:
-        sub_queries: List of sub queries
+        A list of sub-queries
     """
-    # Get web search results prior to generating subqueries for improved context around real time data tasks
-    search_retriever = retriever(query)
-    search_results = search_retriever.search()
-
-    max_research_iterations = cfg.max_iterations if cfg.max_iterations else 1
-    response = await create_chat_completion(
-        model=cfg.smart_llm_model,
-        messages=[
-            {"role": "system", "content": f"{agent_role_prompt}"},
-            {
-                "role": "user",
-                "content": generate_search_queries_prompt(
-                    query,
-                    parent_query,
-                    report_type,
-                    max_iterations=max_research_iterations,
-                    context=search_results
-                ),
-            },
-        ],
-        temperature=0.1,
-        llm_provider=cfg.smart_llm_provider,
-        llm_kwargs=cfg.llm_kwargs,
-        cost_callback=cost_callback,
+
+    sub_queries = await generate_sub_queries(
+        query,
+        parent_query,
+        report_type,
+        search_results,
+        cfg,
+        cost_callback
     )
 
-    sub_queries = json_repair.loads(response)
-
     return sub_queries

gpt_researcher/config/config.py

@@ -48,6 +48,7 @@ def _set_embedding_attributes(self) -> None:
     def _set_llm_attributes(self) -> None:
         self.fast_llm_provider, self.fast_llm_model = self.parse_llm(self.fast_llm)
         self.smart_llm_provider, self.smart_llm_model = self.parse_llm(self.smart_llm)
+        self.strategic_llm_provider, self.strategic_llm_model = self.parse_llm(self.strategic_llm)
 
     def _handle_deprecated_attributes(self) -> None:
         if os.getenv("EMBEDDING_PROVIDER") is not None:
@@ -110,9 +111,10 @@ def load_config(cls, config_path: str | None) -> Dict[str, Any]:
 
         # config_path = os.path.join(cls.CONFIG_DIR, config_path)
         if not os.path.exists(config_path):
-            print(f"Warning: Configuration not found at '{config_path}'. Using default configuration.")
-            if not config_path.endswith(".json"):
-                print(f"Do you mean '{config_path}.json'?")
+            if config_path:
+                print(f"Warning: Configuration not found at '{config_path}'. Using default configuration.")
+                if not config_path.endswith(".json"):
+                    print(f"Do you mean '{config_path}.json'?")
             return DEFAULT_CONFIG
 
         with open(config_path, "r") as f:

gpt_researcher/config/variables/base.py

@@ -8,6 +8,7 @@ class BaseConfig(TypedDict):
     SIMILARITY_THRESHOLD: float
     FAST_LLM: str
     SMART_LLM: str
+    STRATEGIC_LLM: str
     FAST_TOKEN_LIMIT: int
     SMART_TOKEN_LIMIT: int
     BROWSE_CHUNK_MAX_LENGTH: int

gpt_researcher/config/variables/default.py

@@ -6,6 +6,7 @@
     "SIMILARITY_THRESHOLD": 0.42,
     "FAST_LLM": "openai:gpt-4o-mini",
     "SMART_LLM": "openai:gpt-4o-2024-08-06",
+    "STRATEGIC_LLM": "openai:o1-preview",
     "FAST_TOKEN_LIMIT": 2000,
     "SMART_TOKEN_LIMIT": 4000,
     "BROWSE_CHUNK_MAX_LENGTH": 8192,

gpt_researcher/prompts.py

@@ -1,15 +1,16 @@
 import warnings
 from datetime import date, datetime, timezone
 
-from gpt_researcher.utils.enum import ReportSource, ReportType, Tone
+from .utils.enum import ReportSource, ReportType, Tone
+from typing import List, Dict, Any
 
 
 def generate_search_queries_prompt(
     question: str,
     parent_query: str,
     report_type: str,
     max_iterations: int = 3,
-    context: str = ""
+    context: List[Dict[str, Any]] = [],
 ):
     """Generates the search queries prompt for the given question.
     Args:
@@ -31,6 +32,7 @@ def generate_search_queries_prompt(
         task = question
 
     context_prompt = f"""
+You are a seasoned research assistant tasked with generating search queries to find relevant information for the following task: "{task}".
 Context: {context}
 
 Use this context to inform and refine your search queries. The context provides real-time web information that can help you generate more specific and relevant queries. Consider any current events, recent developments, or specific details mentioned in the context that could enhance the search queries.

gpt_researcher/skills/researcher.py

@@ -4,7 +4,7 @@
 from typing import Dict, Optional
 
 from ..actions.utils import stream_output
-from ..actions.query_processing import get_sub_queries
+from ..actions.query_processing import plan_research_outline, get_search_results
 from ..document import DocumentLoader, LangChainDocumentLoader
 from ..utils.enum import ReportSource, ReportType, Tone
 
@@ -111,7 +111,7 @@ async def __get_context_by_vectorstore(self, query, filter: Optional[dict] = Non
         """
         context = []
         # Generate Sub-Queries including original query
-        sub_queries = await self.get_sub_queries(query)
+        sub_queries = await self.plan_research(query)
         # If this is not part of a sub researcher, add original query to research for better results
         if self.researcher.report_type != "subtopic_report":
             sub_queries.append(query)
@@ -143,7 +143,7 @@ async def __get_context_by_search(self, query, scraped_data: list = []):
         """
         context = []
         # Generate Sub-Queries including original query
-        sub_queries = await self.get_sub_queries(query)
+        sub_queries = await self.plan_research(query)
         # If this is not part of a sub researcher, add original query to research for better results
         if self.researcher.report_type != "subtopic_report":
             sub_queries.append(query)
@@ -305,17 +305,26 @@ async def __scrape_data_by_query(self, sub_query):
 
         return scraped_content
 
-    async def get_sub_queries(self, query):
+    async def plan_research(self, query):
         await stream_output(
             "logs",
             "planning_research",
-            f"🌐 Browsing the web and planning research for query: {query}...",
+            f"🌐 Browsing the web to learn more about the task: {query}...",
             self.researcher.websocket,
         )
 
-        return await get_sub_queries(
+        search_results = await get_search_results(query, self.researcher.retrievers[0])
+
+        await stream_output(
+            "logs",
+            "planning_research",
+            f"🤔 Planning the research strategy and subtasks...",
+            self.researcher.websocket,
+        )
+
+        return await plan_research_outline(
             query=query,
-            retriever=self.researcher.retrievers[0],
+            search_results=search_results,
             agent_role_prompt=self.researcher.role,
             cfg=self.researcher.cfg,
             parent_query=self.researcher.parent_query,

gpt_researcher/utils/llm.py

@@ -22,7 +22,7 @@ def get_llm(llm_provider, **kwargs):
 async def create_chat_completion(
         messages: list,  # type: ignore
         model: Optional[str] = None,
-        temperature: float = 0.4,
+        temperature: Optional[float] = 0.4,
         max_tokens: Optional[int] = 4000,
         llm_provider: Optional[str] = None,
         stream: Optional[bool] = False,

requirements.txt

@@ -9,9 +9,9 @@ pydantic
 fastapi
 python-multipart
 markdown
-langchain>=0.2,<0.3
-langchain_community>=0.2,<0.3
-langchain-openai>=0.1,<0.2
+langchain>=0.2,<0.4
+langchain_community>=0.2,<0.4
+langchain-openai>=0.1,<0.4
 langgraph
 tiktoken
 gpt-researcher