Minor tweaks to retrieval skill and file ingestion (#9093)

* Update default params and handle missing embedding args

The default chunk size and overlap values in file_handler have been increased for better performance. In langchain_handler, checks have been added to use the same provider and model for embeddings if 'embedding_model_args' are not provided in input params. Skill_tool's retriever_config has been renamed to 'config' for clarity.

* Handle invoke errors in langchain_handler

This update modifies the langchain handler to accommodate errors during the invocation process. Specifically, it implements exception handling for the agent_executor's invoke method. If an error occurs, instead of crashing, it will now return the error message. However, if the error doesn't match a specific format, the exception will still be raised.

* Update file_handler.py

* Add logging and refactor default vector store in langchain handler

This change introduces logging for situations where 'vector_store_config' is not present in the `langchain_handler` tool configuration. It also modifies the condition that checks for the absence of this property to add a persisting directory. Furthermore, it refactors the code in `langchain_handler.py` related to the absence of 'embedding_model_args'.

* Remove default vector store from langchain handler

The default vector store setup was removed from the langchain handler. The 'vector_store_config' is no longer automatically assigned, and an alert for this missing parameter is no longer logged. The code was adjusted to operate without the previously used 'mindsdb_path'.

mindsdb/integrations/handlers/file_handler/file_handler.py

@@ -26,8 +26,8 @@
 
 logger = log.getLogger(__name__)
 
-DEFAULT_CHUNK_SIZE = 200
-DEFAULT_CHUNK_OVERLAP = 50
+DEFAULT_CHUNK_SIZE = 500
+DEFAULT_CHUNK_OVERLAP = 250
 
 
 def clean_cell(val):

mindsdb/integrations/handlers/langchain_handler/langchain_handler.py

@@ -212,12 +212,22 @@ def create_agent(self, df: pd.DataFrame, args: Dict=None, pred_args: Dict=None)
 
         # Set up embeddings model if needed.
         if args.get('mode') == 'retrieval':
-            # get args for embeddings model
+
             embeddings_args = args.pop('embedding_model_args', {})
+
+            # no embedding model args provided, use same provider as llm
+            if not embeddings_args:
+                logger.warning("'embedding_model_args' not found in input params, "
+                               "Trying to use the same provider used for llm. "
+                               f"provider: {args['provider']}"
+                               )
+
+                # get args for embeddings model
+                embeddings_args['class'] = args['provider']
+
             # create embeddings model
             pred_args['embeddings_model'] = self._create_embeddings_model(embeddings_args)
             pred_args['llm'] = llm
-            pred_args['mindsdb_path'] = self.engine_storage.folder_get
 
         tools = setup_tools(llm,
                             model_kwargs,
@@ -291,8 +301,13 @@ def run_agent(self, df: pd.DataFrame, agent: AgentExecutor, args: Dict, pred_arg
         def _invoke_agent_executor_with_prompt(agent_executor, prompt):
             if not prompt:
                 return ''
-
-            answer = agent_executor.invoke(prompt)
+            try:
+                answer = agent_executor.invoke(prompt)
+            except Exception as e:
+                answer = str(e)
+                if not answer.startswith("Could not parse LLM output: `"):
+                    raise e
+                answer = {'output': answer.removeprefix("Could not parse LLM output: `").removesuffix("`")}
 
             if 'output' not in answer:
                 # This should never happen unless Langchain changes invoke output format, but just in case.

mindsdb/integrations/handlers/langchain_handler/tools.py

@@ -19,6 +19,9 @@
 from mindsdb.integrations.utilities.rag.rag_pipeline_builder import RAG
 from mindsdb.integrations.utilities.rag.settings import RAGPipelineModel
 from mindsdb.interfaces.skills.skill_tool import skill_tool, SkillType
+from mindsdb.utilities import log
+
+logger = log.getLogger(__name__)
 
 # Individual tools
 # Note: all tools are defined in a closure to pass required args (apart from LLM input) through it, as custom tools don't allow custom field assignment.  # noqa
@@ -165,17 +168,11 @@ def _build_retrieval_tool(tool: dict, pred_args: dict):
 
     tools_config = tool['config']
 
-    mindsdb_path = pred_args['mindsdb_path']
-
     # we update the config with the pred_args to allow for custom config
     tools_config.update(pred_args)
 
     rag_params = _get_rag_params(tools_config)
 
-    if 'vector_store_config' not in rag_params:
-
-        rag_params['vector_store_config'] = {'persist_directory': mindsdb_path('persisted_chroma')}
-
     rag_config = RAGPipelineModel(**rag_params)
 
     # build retriever
@@ -188,6 +185,7 @@ def _build_retrieval_tool(tool: dict, pred_args: dict):
         description=tool['description']
     )
 
+
 def langchain_tool_from_skill(skill, pred_args):
     # Makes Langchain compatible tools from a skill
     tool = skill_tool.get_tool_from_skill(skill)

mindsdb/interfaces/skills/skill_tool.py

@@ -84,7 +84,7 @@ def _make_retrieval_tools(self, skill: db.Skills) -> dict:
         params = skill.params
         return dict(
             name=params.get('name', skill.name),
-            config=params.get('retriever_config', {}),
+            config=params.get('config', {}),
             description=f'You must use this tool to get more context or information '
                         f'to answer a question about {params["description"]}. '
                         f'The input should be the exact question the user is asking.',