langroid 0.59.8__tar.gz → 0.59.10__tar.gz

{langroid-0.59.8 → langroid-0.59.10}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: langroid
-Version: 0.59.8
+Version: 0.59.10
 Summary: Harness LLMs with Multi-Agent Programming
 Author-email: Prasad Chalasani <pchalasani@gmail.com>
 License: MIT

{langroid-0.59.8 → langroid-0.59.10}/langroid/agent/special/relevance_extractor_agent.py

@@ -14,7 +14,6 @@ from langroid.agent.chat_document import ChatDocument
 from langroid.agent.tools.segment_extract_tool import SegmentExtractTool
 from langroid.language_models.base import LLMConfig
 from langroid.language_models.openai_gpt import OpenAIGPTConfig
-from langroid.mytypes import Entity
 from langroid.parsing.utils import extract_numbered_segments, number_segments
 from langroid.utils.constants import DONE, NO_ANSWER
 
@@ -26,12 +25,19 @@ class RelevanceExtractorAgentConfig(ChatAgentConfig):
     llm: LLMConfig | None = OpenAIGPTConfig()
     segment_length: int = 1  # number of sentences per segment
     query: str = ""  # query for relevance extraction
+    handle_llm_no_tool: str = """
+    You FORGOT to use the `extract_segments` tool!
+    Remember that your response MUST be a JSON-formatted string
+    starting with `{"request": "extract_segments", ...}`
+    """
     system_message: str = """
     The user will give you a PASSAGE containing segments numbered as  
     <#1#>, <#2#>, <#3#>, etc.,
     followed by a QUERY. Extract ONLY the segment-numbers from 
     the PASSAGE that are RELEVANT to the QUERY.
     Present the extracted segment-numbers using the `extract_segments` tool/function.
+    Note that your response MUST be a JSON-formatted string 
+    starting with `{"request": "extract_segments", ...}`
     """
 
 
@@ -63,13 +69,22 @@ class RelevanceExtractorAgent(ChatAgent):
         self.numbered_passage = number_segments(message_str, self.config.segment_length)
         # compose prompt
         prompt = f"""
+        <Instructions>
+        Given the PASSAGE below with NUMBERED segments, and the QUERY,
+        extract ONLY the segment-numbers that are RELEVANT to the QUERY,
+        and present them using the `extract_segments` tool/function,
+        i.e. your response MUST be a JSON-formatted string starting with
+        `{{"request": "extract_segments", ...}}`
+        </Instructions>
+        
         PASSAGE:
         {self.numbered_passage}
         
         QUERY: {self.config.query}
         """
         # send to LLM
-        return super().llm_response(prompt)
+        response = super().llm_response(prompt)
+        return response
 
     @no_type_check
     async def llm_response_async(
@@ -99,7 +114,8 @@ class RelevanceExtractorAgent(ChatAgent):
         QUERY: {self.config.query}
         """
         # send to LLM
-        return await super().llm_response_async(prompt)
+        response = await super().llm_response_async(prompt)
+        return response
 
     def extract_segments(self, msg: SegmentExtractTool) -> str:
         """Method to handle a segmentExtractTool message from LLM"""
@@ -116,12 +132,3 @@ class RelevanceExtractorAgent(ChatAgent):
             return DONE + " " + NO_ANSWER
         # this response ends the task by saying DONE
         return DONE + " " + extracts
-
-    def handle_message_fallback(
-        self, msg: str | ChatDocument
-    ) -> str | ChatDocument | None:
-        """Handle case where LLM forgets to use SegmentExtractTool"""
-        if isinstance(msg, ChatDocument) and msg.metadata.sender == Entity.LLM:
-            return DONE + " " + NO_ANSWER
-        else:
-            return None

{langroid-0.59.8 → langroid-0.59.10}/langroid/embedding_models/models.py

@@ -463,12 +463,12 @@ class LlamaCppServerEmbeddings(EmbeddingModel):
         response = requests.post(self.embedding_url, json=data)
 
         if response.status_code == 200:
-            embeddings = response.json()["embedding"]
+            embeddings = self._extract_embedding(response.json())
             if not (
                 isinstance(embeddings, list) and isinstance(embeddings[0], (int, float))
             ):
                 raise ValueError(
-                    """Embedding endpoint has not returned the correct format. 
+                    """Embedding endpoint has not returned the correct format.
                    Is the URL correct?
                 """
                 )
@@ -480,6 +480,71 @@ class LlamaCppServerEmbeddings(EmbeddingModel):
                 "Failed to connect to embedding provider",
             )
 
+    def _extract_embedding(
+        self, response_json: dict[str, Any] | list[Any]
+    ) -> List[int | float]:
+        """
+        Extract embedding vector from llama.cpp response.
+
+        Handles multiple response formats:
+        1. Native /embedding: {"embedding": [floats]}
+        2. Array format: [{"embedding": [floats]}]
+        3. Double-nested: [{"embedding": [[floats]]}]
+        4. OpenAI /v1/embeddings: {"data": [{"embedding": [floats]}]}
+        5. Nested in dict: {"embedding": [[floats]]}
+
+        Args:
+            response_json: The JSON response from llama.cpp server
+
+        Returns:
+            List of floats representing the embedding vector
+
+        Raises:
+            ValueError: If response format is not recognized
+        """
+        import json
+
+        # Try native format first: {"embedding": [floats]}
+        if isinstance(response_json, dict) and "embedding" in response_json:
+            embeddings = response_json["embedding"]
+            # Check if it's [floats]
+            if isinstance(embeddings, list) and len(embeddings) > 0:
+                if isinstance(embeddings[0], (int, float)):
+                    return embeddings
+                # Might be nested: {"embedding": [[floats]]}
+                if isinstance(embeddings[0], list) and len(embeddings[0]) > 0:
+                    if isinstance(embeddings[0][0], (int, float)):
+                        return embeddings[0]
+
+        # Try OpenAI format: {"data": [{"embedding": [floats]}]}
+        if isinstance(response_json, dict) and "data" in response_json:
+            data = response_json["data"]
+            if isinstance(data, list) and len(data) > 0:
+                if isinstance(data[0], dict) and "embedding" in data[0]:
+                    embeddings = data[0]["embedding"]
+                    if isinstance(embeddings, list) and len(embeddings) > 0:
+                        if isinstance(embeddings[0], (int, float)):
+                            return embeddings
+
+        # Try array format: [{"embedding": [floats]}] or [{"embedding": [[floats]]}]
+        if isinstance(response_json, list) and len(response_json) > 0:
+            first_item = response_json[0]
+            if isinstance(first_item, dict) and "embedding" in first_item:
+                embeddings = first_item["embedding"]
+                # Check if it's [floats]
+                if isinstance(embeddings, list) and len(embeddings) > 0:
+                    if isinstance(embeddings[0], (int, float)):
+                        return embeddings
+                    # Check if it's [[floats]]
+                    if isinstance(embeddings[0], list) and len(embeddings[0]) > 0:
+                        if isinstance(embeddings[0][0], (int, float)):
+                            return embeddings[0]
+
+        raise ValueError(
+            f"Unsupported embedding response format from {self.embedding_url}. "
+            f"Response: {json.dumps(response_json)[:500]}"
+        )
+
     def embedding_fn(self) -> Callable[[List[str]], Embeddings]:
         return EmbeddingFunctionCallable(self, self.config.batch_size)
 

{langroid-0.59.8 → langroid-0.59.10}/langroid/parsing/parse_json.py

@@ -142,8 +142,14 @@ def top_level_json_field(s: str, f: str) -> Any:
         return ""
     for j in jsons:
         json_data = json.loads(j)
-        if f in json_data:
-            return json_data[f]
+        if isinstance(json_data, dict):
+            if f in json_data:
+                return json_data[f]
+        elif isinstance(json_data, list):
+            # Some responses wrap candidate JSON objects in a list; scan them.
+            for item in json_data:
+                if isinstance(item, dict) and f in item:
+                    return item[f]
 
     return ""
 

{langroid-0.59.8 → langroid-0.59.10}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "langroid"
-version = "0.59.8"
+version = "0.59.10"
 authors = [
     {name = "Prasad Chalasani", email = "pchalasani@gmail.com"},
 ]