Merge pull request #9075 from mindsdb/staging

Release 24.4.3.0

docs/integrations/ai-engines/mindsdb_inference.mdx

@@ -0,0 +1,74 @@
+---
+title: MindsDB Inference Endpoints
+sidebarTitle: MindsDB Inference Endpoints
+---
+
+This documentation describes the integration of MindsDB with [MindsDB Inference Endpoints](https://mindsdb-docs.hashnode.space/), a cloud service that simplifies the way developers interact with cutting-edge LLMs through a universal API.
+
+## Prerequisites
+
+Before proceeding, ensure the following prerequisites are met:
+
+1. Install MindsDB [locally via Docker](https://docs.mindsdb.com/setup/self-hosted/docker) or use [MindsDB Cloud](https://cloud.mindsdb.com/).
+2. To use MindsDB Inference Endpoints within MindsDB, install the required dependencies following [this instruction](/setup/self-hosted/docker#install-dependencies).
+3. Obtain the MindsDB Inference API key required to deploy and use MindsDB Inference Endpoints models within MindsDB. Follow the [instructions for obtaining the API key](https://mindsdb-docs.hashnode.space/docs/authentication).
+
+## Setup
+
+Create an AI engine from the [MindsDB Inference Endpoints handler](https://github.com/mindsdb/mindsdb/tree/staging/mindsdb/integrations/handlers/mindsdb_inference_handler).
+
+```sql
+CREATE ML_ENGINE mindsdb_serve
+FROM mindsdb_inference
+USING
+      mindsdb_inference_api_key = 'api-key-value'
+```
+
+Create a model using `mindsdb_serve` as an engine.
+
+```sql
+CREATE MODEL mindsdb_inference_model
+PREDICT answer
+USING
+      engine = 'mindsdb_serve',   -- engine name as created via CREATE ML_ENGINE
+      model_name = 'model-name',              -- choose one of available models
+      prompt_teplate = 'prompt-to-the-model'; -- prompt message to be completed by the model
+      mode = 'mode_name' -- choose one of the available modes: 'conversational', 'conversational-full'
+```
+
+## Usage
+
+The following usage examples utilize `mindsdb_serve` to create a model with the `CREATE MODEL` statement.
+
+Classify text sentiment using the Mistral 7B model.
+
+```sql
+CREATE MODEL mindsdb_inference_model
+PREDICT sentiment
+USING
+   engine = 'mindsdb_serve',
+   model_name = 'mistral-7b',
+   prompt_template = 'Classify the sentiment of the following text as one of `positive`, `neutral` or `negative`: {{text}}';
+```
+
+Query the model to get predictions.
+
+```sql
+SELECT text, sentiment
+FROM mindsdb_inference_model
+WHERE text = 'I love machine learning!';
+```
+
+Here is the output:
+
+```sql
++--------------------------+-----------+
+| text                     | sentiment |
++--------------------------+-----------+
+| I love machine learning! | positive  |
++--------------------------+-----------+
+```
+
+## Supported Models
+
+For an overview of the models supported, visit the [following docs](https://docs.mdb.ai/). This list will help you quickly identify the right models for your needs.

docs/mindsdb_sql/agents/knowledge-bases.mdx

@@ -27,6 +27,18 @@ CREATE KNOWLEDGE_BASE my_kb
    [storage = vector_database.storage_table;]]
 ```
 
+### Managing input columns
+
+Knowledge base can accept optional columns parameters to define where id, content and metadata columns are located:
+```sql
+CREATE KNOWLEDGE_BASE my_kb
+USING
+   metadata_columns = ['date', 'creator'],  -- optional, if not set: no metadata columns
+   content_columns = ['review'],  -- optional, if not set: all columns is content
+   id_column='index'   -- optional, default: id
+```
+
+
 ## Step by step guide
 
 

docs/mint.json

@@ -977,6 +977,7 @@
             "integrations/ai-engines/langchain",
             "integrations/ai-engines/llamaindex",
             "integrations/ai-engines/monkeylearn",
+            "integrations/ai-engines/mindsdb_inference",
             "integrations/ai-engines/ollama",
             "integrations/ai-engines/openai",
             "integrations/ai-engines/replicate-llm",

docs/sql/tutorials/text-sentiment-hf.mdx

@@ -3,7 +3,7 @@ title: Predict Text Sentiment with Hugging Face and MindsDB
 sidebarTitle: Predict Text Sentiment
 ---
 
-In this tutorial, we'll use a model from the Hugging Fae hub to predict text sentiment.
+In this tutorial, we'll use a model from the Hugging Faсe hub to predict text sentiment.
 
 ## Connect a database
 
@@ -44,7 +44,7 @@ USING
    labels=['negative','neutral','positive'];
 ```
 
-To create a model in MindsDB, we use the `CREATE MODEL` statement. Next, we define the target column using the PREDICT clause. Finally, we specify all required parameters in the `USING` clause.
+To create a model in MindsDB, we use the `CREATE MODEL` statement. Next, we define the target column using the `PREDICT` clause. Finally, we specify all required parameters in the `USING` clause.
 
 Once the above query is executed, we can check the status of the creation process:
 

mindsdb/__about__.py

@@ -1,6 +1,6 @@
 __title__ = 'MindsDB'
 __package_name__ = 'mindsdb'
-__version__ = '24.4.2.1'
+__version__ = '24.4.3.0'
 __description__ = "MindsDB's AI SQL Server enables developers to build AI tools that need access to real-time data to perform their tasks"
 __email__ = "jorge@mindsdb.com"
 __author__ = 'MindsDB Inc'

mindsdb/api/executor/datahub/datanodes/information_schema_datanode.py

@@ -315,7 +315,7 @@ class InformationSchemaDataNode(DataNode):
             "IS_RUNNING",
             "LAST_ERROR",
         ],
-        "KNOWLEDGE_BASES": ["NAME", "PROJECT", "MODEL", "STORAGE"],
+        "KNOWLEDGE_BASES": ["NAME", "PROJECT", "MODEL", "STORAGE", "PARAMS"],
         "SKILLS": ["NAME", "PROJECT", "TYPE", "PARAMS"],
         "AGENTS": [
             "NAME",
@@ -672,7 +672,7 @@ def _get_knowledge_bases(self, query: ASTNode = None):
 
         columns = self.information_schema['KNOWLEDGE_BASES']
 
-        # columns: NAME, PROJECT, MODEL, STORAGE
+        # columns: NAME, PROJECT, MODEL, STORAGE, PARAMS
         data = []
 
         for kb in kb_list:
@@ -684,7 +684,8 @@ def _get_knowledge_bases(self, query: ASTNode = None):
                 kb.name,
                 project_name,
                 embedding_model.name if embedding_model is not None else None,
-                vector_database_name + '.' + kb.vector_database_table
+                vector_database_name + '.' + kb.vector_database_table,
+                to_json(kb.params),
             ))
 
         return pd.DataFrame(data, columns=columns)

mindsdb/integrations/handlers/chromadb_handler/tests/test_chromadb_handler.py

@@ -6,7 +6,6 @@
 
 import pandas as pd
 import pytest
-from mindsdb_sql import parse_sql
 
 from tests.unit.executor_test_base import BaseExecutorTest
 
@@ -19,15 +18,6 @@
 
 @pytest.mark.skipif(not CHROMA_DB_INSTALLED, reason="chroma_db is not installed")
 class TestChromaDBHandler(BaseExecutorTest):
-    def run_sql(self, sql):
-        ret = self.command_executor.execute_command(parse_sql(sql, dialect="mindsdb"))
-
-        assert ret.error_code is None
-        if ret.data is not None:
-            columns = [
-                col.alias if col.alias is not None else col.name for col in ret.columns
-            ]
-            return pd.DataFrame(ret.data, columns=columns)
 
     @pytest.fixture(autouse=True, scope="function")
     def setup_method(self):

mindsdb/integrations/handlers/langchain_handler/constants.py

@@ -87,7 +87,7 @@
 DEFAULT_AGENT_TIMEOUT_SECONDS = 300
 # These should require no additional arguments.
 DEFAULT_AGENT_TOOLS = []  
-DEFAULT_AGENT_TYPE = AgentType.CONVERSATIONAL_REACT_DESCRIPTION
+DEFAULT_AGENT_TYPE = AgentType.OPENAI_FUNCTIONS
 DEFAULT_MAX_ITERATIONS = 10
 DEFAULT_MAX_TOKENS = 2048
 DEFAULT_MODEL_NAME = 'gpt-4-0125-preview'

mindsdb/integrations/handlers/langchain_handler/langchain_handler.py

@@ -27,13 +27,14 @@
     DEFAULT_ASSISTANT_COLUMN
 )
 from mindsdb.integrations.handlers.langchain_handler.log_callback_handler import LogCallbackHandler
-from mindsdb.integrations.handlers.langchain_handler.mindsdb_database_agent import MindsDBSQL
+from mindsdb.integrations.utilities.rag.settings import DEFAULT_RAG_PROMPT_TEMPLATE
 from mindsdb.integrations.handlers.langchain_handler.tools import setup_tools
 from mindsdb.integrations.handlers.openai_handler.constants import CHAT_MODELS as OPEN_AI_CHAT_MODELS
 from mindsdb.integrations.libs.base import BaseMLEngine
 from mindsdb.integrations.libs.llm.utils import get_llm_config
 from mindsdb.integrations.utilities.handler_utils import get_api_key
 from mindsdb.interfaces.storage.model_fs import HandlerStorage, ModelStorage
+from mindsdb.integrations.handlers.langchain_embedding_handler.langchain_embedding_handler import construct_model_from_args
 from mindsdb.utilities import log
 from mindsdb.utilities.context_executor import ContextThreadPoolExecutor
 
@@ -76,7 +77,7 @@ def __init__(
         if self.log_callback_handler is None:
             self.log_callback_handler = LogCallbackHandler(logger)
 
-    def _get_provider(self, args: Dict) -> str:
+    def _get_llm_provider(self, args: Dict) -> str:
         if 'provider' in args:
             return args['provider']
         if args['model_name'] in ANTHROPIC_CHAT_MODELS:
@@ -87,6 +88,14 @@ def _get_provider(self, args: Dict) -> str:
             return 'ollama'
         raise ValueError(f"Invalid model name. Please define provider")
 
+    def _get_embedding_model_provider(self, args: Dict) -> str:
+        if 'embedding_model_provider' in args:
+            return args['embedding_model_provider']
+        if 'embedding_model_provider' not in args:
+            logger.warning('No embedding model provider specified. trying to use llm provider.')
+            return args.get('embedding_model_provider', self._get_llm_provider(args))
+        raise ValueError(f"Invalid model name. Please define provider")
+
     def _get_chat_model_params(self, args: Dict, pred_args: Dict) -> Dict:
         model_config = args.copy()
         # Override with prediction args.
@@ -95,8 +104,8 @@ def _get_chat_model_params(self, args: Dict, pred_args: Dict) -> Dict:
         model_config['api_keys'] = {
             p: get_api_key(p, args, self.engine_storage, strict=False) for p in SUPPORTED_PROVIDERS
         }
-        llm_config = get_llm_config(args.get('provider', self._get_provider(args)), model_config)
-        config_dict = llm_config.model_dump()
+        llm_config = get_llm_config(args.get('provider', self._get_llm_provider(args)), model_config)
+        config_dict = llm_config.dict()
         config_dict = {k: v for k, v in config_dict.items() if v is not None}
         return config_dict
 
@@ -131,6 +140,9 @@ def _create_chat_model(self, args: Dict, pred_args: Dict):
             return ChatOllama(**model_kwargs)
         raise ValueError(f'Unknown provider: {args["provider"]}')
 
+    def _create_embeddings_model(self, args: Dict):
+        return construct_model_from_args(args)
+
     def _handle_parsing_errors(self, error: Exception) -> str:
         response = str(error)
         if not response.startswith(_PARSING_ERROR_PREFIX):
@@ -152,7 +164,13 @@ def create(self, target: str, args: Dict = None, **kwargs):
         args = args['using']
         args['target'] = target
         args['model_name'] = args.get('model_name', DEFAULT_MODEL_NAME)
-        args['provider'] = args.get('provider', self._get_provider(args))
+        args['provider'] = args.get('provider', self._get_llm_provider(args))
+        args['embedding_model_provider'] = args.get('embedding_model', self._get_embedding_model_provider(args))
+        if args.get('mode') == 'retrieval':
+            # use default prompt template for retrieval i.e. RAG if not provided
+            if "prompt_template" not in args:
+                args["prompt_template"] = DEFAULT_RAG_PROMPT_TEMPLATE
+
         self.model_storage.json_set('args', args)
 
     @staticmethod
@@ -162,7 +180,8 @@ def create_validation(_, args: Dict=None, **kwargs):
         else:
             args = args['using']
         if 'prompt_template' not in args:
-            raise ValueError('Please provide a `prompt_template` for this engine.')
+            if not args.get('mode') == 'retrieval':
+                raise ValueError('Please provide a `prompt_template` for this engine.')
 
     def predict(self, df: pd.DataFrame, args: Dict=None) -> pd.DataFrame:
         """
@@ -174,7 +193,8 @@ def predict(self, df: pd.DataFrame, args: Dict=None) -> pd.DataFrame:
         if 'prompt_template' not in args and 'prompt_template' not in pred_args:
             raise ValueError(f"This model expects a `prompt_template`, please provide one.")
         # Back compatibility for old models
-        args['provider'] = args.get('provider', self._get_provider(args))
+        args['provider'] = args.get('provider', self._get_llm_provider(args))
+        args['embedding_model_provider'] = args.get('embedding_model', self._get_embedding_model_provider(args))
 
         df = df.reset_index(drop=True)
         agent = self.create_agent(df, args, pred_args)
@@ -189,6 +209,16 @@ def create_agent(self, df: pd.DataFrame, args: Dict=None, pred_args: Dict=None)
         # Set up tools.
         model_kwargs = self._get_chat_model_params(args, pred_args)
         llm = self._create_chat_model(args, pred_args)
+
+        # Set up embeddings model if needed.
+        if args.get('mode') == 'retrieval':
+            # get args for embeddings model
+            embeddings_args = args.pop('embedding_model_args', {})
+            # 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,
                             pred_args,
@@ -228,7 +258,7 @@ def create_agent(self, df: pd.DataFrame, args: Dict=None, pred_args: Dict=None)
             max_execution_time=pred_args.get('timeout_seconds', args.get('timeout_seconds', DEFAULT_AGENT_TIMEOUT_SECONDS)),
             max_iterations=pred_args.get('max_iterations', args.get('max_iterations', DEFAULT_MAX_ITERATIONS)),
             memory=memory,
-            verbose=pred_args.get('verbose', args.get('verbose', False))
+            verbose=pred_args.get('verbose', args.get('verbose', True))
         )
         return agent_executor
 

mindsdb/integrations/handlers/langchain_handler/log_callback_handler.py

@@ -13,6 +13,7 @@ class LogCallbackHandler(BaseCallbackHandler):
     def __init__(self, logger: logging.Logger):
         logger.setLevel('DEBUG')
         self.logger = logger
+        self._num_running_chains = 0
 
     def on_llm_start(
         self, serialized: Dict[str, Any], prompts: List[str], **kwargs: Any

mindsdb/integrations/handlers/langchain_handler/requirements.txt

@@ -4,5 +4,6 @@ tiktoken >= 0.3.0
 anthropic==0.3.5
 langchain-experimental  # Tools
 langfuse  # Tracing
-litellm==1.23.2
+litellm==1.35.0
 -r mindsdb/integrations/handlers/openai_handler/requirements.txt
+-r mindsdb/integrations/handlers/langchain_embedding_handler/requirements.txt