import os import shutil from collections import defaultdict from datetime import datetime from typing import ( TYPE_CHECKING, Any, Callable, Dict, List, Optional, Sequence, Tuple, TypedDict, Union, ) from llama_index.callbacks.base_handler import BaseCallbackHandler from llama_index.callbacks.schema import ( TIMESTAMP_FORMAT, CBEvent, CBEventType, EventPayload, ) from llama_index.callbacks.token_counting import get_llm_token_counts from llama_index.utilities.token_counting import TokenCounter from llama_index.utils import get_tokenizer if TYPE_CHECKING: from wandb import Settings as WBSettings from wandb.sdk.data_types import trace_tree from llama_index.indices import ( ComposableGraph, GPTEmptyIndex, GPTKeywordTableIndex, GPTRAKEKeywordTableIndex, GPTSimpleKeywordTableIndex, GPTSQLStructStoreIndex, GPTTreeIndex, GPTVectorStoreIndex, SummaryIndex, ) from llama_index.storage.storage_context import StorageContext IndexType = Union[ ComposableGraph, GPTKeywordTableIndex, GPTSimpleKeywordTableIndex, GPTRAKEKeywordTableIndex, SummaryIndex, GPTEmptyIndex, GPTTreeIndex, GPTVectorStoreIndex, GPTSQLStructStoreIndex, ] # remove this class class WandbRunArgs(TypedDict): job_type: Optional[str] dir: Optional[str] config: Union[Dict, str, None] project: Optional[str] entity: Optional[str] reinit: Optional[bool] tags: Optional[Sequence] group: Optional[str] name: Optional[str] notes: Optional[str] magic: Optional[Union[dict, str, bool]] config_exclude_keys: Optional[List[str]] config_include_keys: Optional[List[str]] anonymous: Optional[str] mode: Optional[str] allow_val_change: Optional[bool] resume: Optional[Union[bool, str]] force: Optional[bool] tensorboard: Optional[bool] sync_tensorboard: Optional[bool] monitor_gym: Optional[bool] save_code: Optional[bool] id: Optional[str] settings: Union["WBSettings", Dict[str, Any], None] class WandbCallbackHandler(BaseCallbackHandler): """Callback handler that logs events to wandb. NOTE: this is a beta feature. The usage within our codebase, and the interface may change. Use the `WandbCallbackHandler` to log trace events to wandb. This handler is useful for debugging and visualizing the trace events. It captures the payload of the events and logs them to wandb. The handler also tracks the start and end of events. This is particularly useful for debugging your LLM calls. The `WandbCallbackHandler` can also be used to log the indices and graphs to wandb using the `persist_index` method. This will save the indexes as artifacts in wandb. The `load_storage_context` method can be used to load the indexes from wandb artifacts. This method will return a `StorageContext` object that can be used to build the index, using `load_index_from_storage`, `load_indices_from_storage` or `load_graph_from_storage` functions. Args: event_starts_to_ignore (Optional[List[CBEventType]]): list of event types to ignore when tracking event starts. event_ends_to_ignore (Optional[List[CBEventType]]): list of event types to ignore when tracking event ends. """ def __init__( self, run_args: Optional[WandbRunArgs] = None, tokenizer: Optional[Callable[[str], List]] = None, event_starts_to_ignore: Optional[List[CBEventType]] = None, event_ends_to_ignore: Optional[List[CBEventType]] = None, ) -> None: try: import wandb from wandb.sdk.data_types import trace_tree self._wandb = wandb self._trace_tree = trace_tree except ImportError: raise ImportError( "WandbCallbackHandler requires wandb. " "Please install it with `pip install wandb`." ) from llama_index.indices import ( ComposableGraph, GPTEmptyIndex, GPTKeywordTableIndex, GPTRAKEKeywordTableIndex, GPTSimpleKeywordTableIndex, GPTSQLStructStoreIndex, GPTTreeIndex, GPTVectorStoreIndex, SummaryIndex, ) self._IndexType = ( ComposableGraph, GPTKeywordTableIndex, GPTSimpleKeywordTableIndex, GPTRAKEKeywordTableIndex, SummaryIndex, GPTEmptyIndex, GPTTreeIndex, GPTVectorStoreIndex, GPTSQLStructStoreIndex, ) self._run_args = run_args # Check if a W&B run is already initialized; if not, initialize one self._ensure_run(should_print_url=(self._wandb.run is None)) self._event_pairs_by_id: Dict[str, List[CBEvent]] = defaultdict(list) self._cur_trace_id: Optional[str] = None self._trace_map: Dict[str, List[str]] = defaultdict(list) self.tokenizer = tokenizer or get_tokenizer() self._token_counter = TokenCounter(tokenizer=self.tokenizer) event_starts_to_ignore = ( event_starts_to_ignore if event_starts_to_ignore else [] ) event_ends_to_ignore = event_ends_to_ignore if event_ends_to_ignore else [] super().__init__( event_starts_to_ignore=event_starts_to_ignore, event_ends_to_ignore=event_ends_to_ignore, ) def on_event_start( self, event_type: CBEventType, payload: Optional[Dict[str, Any]] = None, event_id: str = "", parent_id: str = "", **kwargs: Any, ) -> str: """Store event start data by event type. Args: event_type (CBEventType): event type to store. payload (Optional[Dict[str, Any]]): payload to store. event_id (str): event id to store. parent_id (str): parent event id. """ event = CBEvent(event_type, payload=payload, id_=event_id) self._event_pairs_by_id[event.id_].append(event) return event.id_ def on_event_end( self, event_type: CBEventType, payload: Optional[Dict[str, Any]] = None, event_id: str = "", **kwargs: Any, ) -> None: """Store event end data by event type. Args: event_type (CBEventType): event type to store. payload (Optional[Dict[str, Any]]): payload to store. event_id (str): event id to store. """ event = CBEvent(event_type, payload=payload, id_=event_id) self._event_pairs_by_id[event.id_].append(event) self._trace_map = defaultdict(list) def start_trace(self, trace_id: Optional[str] = None) -> None: """Launch a trace.""" self._trace_map = defaultdict(list) self._cur_trace_id = trace_id self._start_time = datetime.now() def end_trace( self, trace_id: Optional[str] = None, trace_map: Optional[Dict[str, List[str]]] = None, ) -> None: # Ensure W&B run is initialized self._ensure_run() self._trace_map = trace_map or defaultdict(list) self._end_time = datetime.now() # Log the trace map to wandb # We can control what trace ids we want to log here. self.log_trace_tree() # TODO (ayulockin): Log the LLM token counts to wandb when weave is ready def log_trace_tree(self) -> None: """Log the trace tree to wandb.""" try: child_nodes = self._trace_map["root"] root_span = self._convert_event_pair_to_wb_span( self._event_pairs_by_id[child_nodes[0]], trace_id=self._cur_trace_id if len(child_nodes) > 1 else None, ) if len(child_nodes) == 1: child_nodes = self._trace_map[child_nodes[0]] root_span = self._build_trace_tree(child_nodes, root_span) else: root_span = self._build_trace_tree(child_nodes, root_span) if root_span: root_trace = self._trace_tree.WBTraceTree(root_span) if self._wandb.run: self._wandb.run.log({"trace": root_trace}) self._wandb.termlog("Logged trace tree to W&B.") except Exception as e: print(f"Failed to log trace tree to W&B: {e}") # ignore errors to not break user code def persist_index( self, index: "IndexType", index_name: str, persist_dir: Union[str, None] = None ) -> None: """Upload an index to wandb as an artifact. You can learn more about W&B artifacts here: https://docs.wandb.ai/guides/artifacts. For the `ComposableGraph` index, the root id is stored as artifact metadata. Args: index (IndexType): index to upload. index_name (str): name of the index. This will be used as the artifact name. persist_dir (Union[str, None]): directory to persist the index. If None, a temporary directory will be created and used. """ if persist_dir is None: persist_dir = f"{self._wandb.run.dir}/storage" # type: ignore _default_persist_dir = True if not os.path.exists(persist_dir): os.makedirs(persist_dir) if isinstance(index, self._IndexType): try: index.storage_context.persist(persist_dir) # type: ignore metadata = None # For the `ComposableGraph` index, store the root id as metadata if isinstance(index, self._IndexType[0]): root_id = index.root_id metadata = {"root_id": root_id} self._upload_index_as_wb_artifact(persist_dir, index_name, metadata) except Exception as e: # Silently ignore errors to not break user code self._print_upload_index_fail_message(e) # clear the default storage dir if _default_persist_dir: shutil.rmtree(persist_dir, ignore_errors=True) def load_storage_context( self, artifact_url: str, index_download_dir: Union[str, None] = None ) -> "StorageContext": """Download an index from wandb and return a storage context. Use this storage context to load the index into memory using `load_index_from_storage`, `load_indices_from_storage` or `load_graph_from_storage` functions. Args: artifact_url (str): url of the artifact to download. The artifact url will be of the form: `entity/project/index_name:version` and can be found in the W&B UI. index_download_dir (Union[str, None]): directory to download the index to. """ from llama_index.storage.storage_context import StorageContext artifact = self._wandb.use_artifact(artifact_url, type="storage_context") artifact_dir = artifact.download(root=index_download_dir) return StorageContext.from_defaults(persist_dir=artifact_dir) def _upload_index_as_wb_artifact( self, dir_path: str, artifact_name: str, metadata: Optional[Dict] ) -> None: """Utility function to upload a dir to W&B as an artifact.""" artifact = self._wandb.Artifact(artifact_name, type="storage_context") if metadata: artifact.metadata = metadata artifact.add_dir(dir_path) self._wandb.run.log_artifact(artifact) # type: ignore def _build_trace_tree( self, events: List[str], span: "trace_tree.Span" ) -> "trace_tree.Span": """Build the trace tree from the trace map.""" for child_event in events: child_span = self._convert_event_pair_to_wb_span( self._event_pairs_by_id[child_event] ) child_span = self._build_trace_tree( self._trace_map[child_event], child_span ) span.add_child_span(child_span) return span def _convert_event_pair_to_wb_span( self, event_pair: List[CBEvent], trace_id: Optional[str] = None, ) -> "trace_tree.Span": """Convert a pair of events to a wandb trace tree span.""" start_time_ms, end_time_ms = self._get_time_in_ms(event_pair) if trace_id is None: event_type = event_pair[0].event_type span_kind = self._map_event_type_to_span_kind(event_type) else: event_type = trace_id # type: ignore span_kind = None wb_span = self._trace_tree.Span( name=f"{event_type}", span_kind=span_kind, start_time_ms=start_time_ms, end_time_ms=end_time_ms, ) inputs, outputs, wb_span = self._add_payload_to_span(wb_span, event_pair) wb_span.add_named_result(inputs=inputs, outputs=outputs) # type: ignore return wb_span def _map_event_type_to_span_kind( self, event_type: CBEventType ) -> Union[None, "trace_tree.SpanKind"]: """Map a CBEventType to a wandb trace tree SpanKind.""" if event_type == CBEventType.CHUNKING: span_kind = None elif event_type == CBEventType.NODE_PARSING: span_kind = None elif event_type == CBEventType.EMBEDDING: # TODO: add span kind for EMBEDDING when it's available span_kind = None elif event_type == CBEventType.LLM: span_kind = self._trace_tree.SpanKind.LLM elif event_type == CBEventType.QUERY: span_kind = self._trace_tree.SpanKind.AGENT elif event_type == CBEventType.AGENT_STEP: span_kind = self._trace_tree.SpanKind.AGENT elif event_type == CBEventType.RETRIEVE: span_kind = self._trace_tree.SpanKind.TOOL elif event_type == CBEventType.SYNTHESIZE: span_kind = self._trace_tree.SpanKind.CHAIN elif event_type == CBEventType.TREE: span_kind = self._trace_tree.SpanKind.CHAIN elif event_type == CBEventType.SUB_QUESTION: span_kind = self._trace_tree.SpanKind.CHAIN elif event_type == CBEventType.RERANKING: span_kind = self._trace_tree.SpanKind.CHAIN elif event_type == CBEventType.FUNCTION_CALL: span_kind = self._trace_tree.SpanKind.TOOL else: span_kind = None return span_kind def _add_payload_to_span( self, span: "trace_tree.Span", event_pair: List[CBEvent] ) -> Tuple[Optional[Dict[str, Any]], Optional[Dict[str, Any]], "trace_tree.Span"]: """Add the event's payload to the span.""" assert len(event_pair) == 2 event_type = event_pair[0].event_type inputs = None outputs = None if event_type == CBEventType.NODE_PARSING: # TODO: disabled full detailed inputs/outputs due to UI lag inputs, outputs = self._handle_node_parsing_payload(event_pair) elif event_type == CBEventType.LLM: inputs, outputs, span = self._handle_llm_payload(event_pair, span) elif event_type == CBEventType.QUERY: inputs, outputs = self._handle_query_payload(event_pair) elif event_type == CBEventType.EMBEDDING: inputs, outputs = self._handle_embedding_payload(event_pair) return inputs, outputs, span def _handle_node_parsing_payload( self, event_pair: List[CBEvent] ) -> Tuple[Dict[str, Any], Dict[str, Any]]: """Handle the payload of a NODE_PARSING event.""" inputs = event_pair[0].payload outputs = event_pair[-1].payload if inputs and EventPayload.DOCUMENTS in inputs: documents = inputs.pop(EventPayload.DOCUMENTS) inputs["num_documents"] = len(documents) if outputs and EventPayload.NODES in outputs: nodes = outputs.pop(EventPayload.NODES) outputs["num_nodes"] = len(nodes) return inputs or {}, outputs or {} def _handle_llm_payload( self, event_pair: List[CBEvent], span: "trace_tree.Span" ) -> Tuple[Dict[str, Any], Dict[str, Any], "trace_tree.Span"]: """Handle the payload of a LLM event.""" inputs = event_pair[0].payload outputs = event_pair[-1].payload assert isinstance(inputs, dict) and isinstance(outputs, dict) # Get `original_template` from Prompt if EventPayload.PROMPT in inputs: inputs[EventPayload.PROMPT] = inputs[EventPayload.PROMPT] # Format messages if EventPayload.MESSAGES in inputs: inputs[EventPayload.MESSAGES] = "\n".join( [str(x) for x in inputs[EventPayload.MESSAGES]] ) token_counts = get_llm_token_counts(self._token_counter, outputs) metadata = { "formatted_prompt_tokens_count": token_counts.prompt_token_count, "prediction_tokens_count": token_counts.completion_token_count, "total_tokens_used": token_counts.total_token_count, } span.attributes = metadata # Make `response` part of `outputs` outputs = {EventPayload.RESPONSE: str(outputs[EventPayload.RESPONSE])} return inputs, outputs, span def _handle_query_payload( self, event_pair: List[CBEvent] ) -> Tuple[Optional[Dict[str, Any]], Dict[str, Any]]: """Handle the payload of a QUERY event.""" inputs = event_pair[0].payload outputs = event_pair[-1].payload if outputs: response_obj = outputs[EventPayload.RESPONSE] response = str(outputs[EventPayload.RESPONSE]) if type(response).__name__ == "Response": response = response_obj.response elif type(response).__name__ == "StreamingResponse": response = response_obj.get_response().response else: response = " " outputs = {"response": response} return inputs, outputs def _handle_embedding_payload( self, event_pair: List[CBEvent], ) -> Tuple[Optional[Dict[str, Any]], Dict[str, Any]]: event_pair[0].payload outputs = event_pair[-1].payload chunks = [] if outputs: chunks = outputs.get(EventPayload.CHUNKS, []) return {}, {"num_chunks": len(chunks)} def _get_time_in_ms(self, event_pair: List[CBEvent]) -> Tuple[int, int]: """Get the start and end time of an event pair in milliseconds.""" start_time = datetime.strptime(event_pair[0].time, TIMESTAMP_FORMAT) end_time = datetime.strptime(event_pair[1].time, TIMESTAMP_FORMAT) start_time_in_ms = int( (start_time - datetime(1970, 1, 1)).total_seconds() * 1000 ) end_time_in_ms = int((end_time - datetime(1970, 1, 1)).total_seconds() * 1000) return start_time_in_ms, end_time_in_ms def _ensure_run(self, should_print_url: bool = False) -> None: """Ensures an active W&B run exists. If not, will start a new run with the provided run_args. """ if self._wandb.run is None: # Make a shallow copy of the run args, so we don't modify the original run_args = self._run_args or {} # type: ignore run_args: dict = {**run_args} # type: ignore # Prefer to run in silent mode since W&B has a lot of output # which can be undesirable when dealing with text-based models. if "settings" not in run_args: # type: ignore run_args["settings"] = {"silent": True} # type: ignore # Start the run and add the stream table self._wandb.init(**run_args) self._wandb.run._label(repo="llama_index") # type: ignore if should_print_url: self._print_wandb_init_message( self._wandb.run.settings.run_url # type: ignore ) def _print_wandb_init_message(self, run_url: str) -> None: """Print a message to the terminal when W&B is initialized.""" self._wandb.termlog( f"Streaming LlamaIndex events to W&B at {run_url}\n" "`WandbCallbackHandler` is currently in beta.\n" "Please report any issues to https://github.com/wandb/wandb/issues " "with the tag `llamaindex`." ) def _print_upload_index_fail_message(self, e: Exception) -> None: """Print a message to the terminal when uploading the index fails.""" self._wandb.termlog( f"Failed to upload index to W&B with the following error: {e}\n" ) def finish(self) -> None: """Finish the callback handler.""" self._wandb.finish()