Skip to content

armgit / letta

Public Imported from GitHub
0

Platform for stateful agents: AI with advanced memory that can learn and self-improve over time.

fix: Patch model dump issues on firecrawl (#2864)

Matthew Zhou75edfda3c357
diff --git a/letta/functions/prompts.py b/letta/functions/prompts.py
index f9bac2a0b..85bd99a94 100644
--- a/letta/functions/prompts.py
+++ b/letta/functions/prompts.py
@@ -6,10 +6,9 @@ Guidelines:
1. Extract substantial, lengthy text snippets that directly address the question
2. Preserve important context and details in each snippet - err on the side of including more rather than less
3. Keep thinking very brief (1 short sentence) - focus on WHY the snippet is relevant, not WHAT it says
-4. Include a concise summary of how the overall document relates to the question
-5. Only extract snippets that actually answer or relate to the question - don't force relevance
-6. Be comprehensive - include all relevant information, don't limit the number of snippets
-7. Prioritize longer, information-rich passages over shorter ones"""
+4. Only extract snippets that actually answer or relate to the question - don't force relevance
+5. Be comprehensive - include all relevant information, don't limit the number of snippets
+6. Prioritize longer, information-rich passages over shorter ones"""
def get_firecrawl_search_user_prompt(query: str, question: str, markdown_content: str) -> str:
diff --git a/letta/services/tool_executor/builtin_tool_executor.py b/letta/services/tool_executor/builtin_tool_executor.py
index d8abe2986..859cb8aef 100644
--- a/letta/services/tool_executor/builtin_tool_executor.py
+++ b/letta/services/tool_executor/builtin_tool_executor.py
@@ -31,7 +31,6 @@ class DocumentAnalysis(BaseModel):
"""Analysis of a document's relevance to a search question."""
citations: List[Citation]
- summary: str # Brief summary of how this document relates to the question
class LettaBuiltinToolExecutor(ToolExecutor):
@@ -225,7 +224,7 @@ class LettaBuiltinToolExecutor(ToolExecutor):
for result, analysis in zip(results_with_markdown, analyses):
if isinstance(analysis, Exception) or analysis is None:
logger.error(f"Analysis failed for {result.get('url')}, falling back to raw results")
- return search_result.model_dump_json(exclude_none=True)
+ return str(search_result)
# All analyses succeeded, build processed results
for result, analysis in zip(results_with_markdown, analyses):
@@ -252,7 +251,7 @@ class LettaBuiltinToolExecutor(ToolExecutor):
logger.error(f"Error with OpenAI processing: {e}")
# Return raw search results if OpenAI processing isn't available or fails
- return search_result.model_dump_json(exclude_none=True)
+ return str(search_result)
async def _analyze_document_with_openai(self, client, markdown_content: str, query: str, question: str) -> Optional[DocumentAnalysis]:
"""Use OpenAI to analyze a document and extract relevant passages."""
@@ -283,11 +282,9 @@ class LettaBuiltinToolExecutor(ToolExecutor):
if result.get("analysis") and result["analysis"].get("citations"):
analysis = result["analysis"]
- source["summary"] = analysis.get("summary")
source["citations"] = analysis["citations"]
total_snippets += len(analysis["citations"])
else:
- source["summary"] = "No relevant information found to answer the question"
source["citations"] = []
sources.append(source)