import asyncio import logging import threading import time from collections import defaultdict from collections.abc import Awaitable, Callable from typing import Any, Literal, NamedTuple, TypeVar import tiktoken from google import genai from google.genai import types as genai_types from nanoid import generate as generate_nanoid from openai import AsyncOpenAI from .config import EmbeddingModelConfig, resolve_embedding_model_config, settings logger = logging.getLogger(__name__) _T = TypeVar("_T") async def _emit_embedding_call( *, provider: str, model: str, texts: list[str], input_tokens_estimate: int, fn: Callable[[], Awaitable[_T]], is_final_attempt: bool = True, ) -> _T: """time a single embedding-provider call, emit `embedding.call.completed` on both success and exception, and return the call's result. Errors propagate unchanged — telemetry never bleeds into the caller's control flow. Caller-supplied `texts` is used only for `input_count`; we don't keep the list around for the event to avoid leaking content into telemetry. `is_final_attempt` defaults to True so one-shot callers (`embed`, `simple_batch_embed`) get correct semantics without changes. Retry-loop callers (`_process_batch`) pass the real attempt index so dashboards can distinguish exhausted retries from mid-retry failures. """ start = time.perf_counter() error: BaseException | None = None try: return await fn() except BaseException as exc: error = exc raise finally: if error is None: outcome: Literal["success", "error", "cancelled"] = "success" elif isinstance(error, asyncio.CancelledError): outcome = "cancelled" else: outcome = "error" _publish_embedding_event( provider=provider, model=model, input_count=len(texts), input_tokens_estimate=input_tokens_estimate, duration_ms=(time.perf_counter() - start) * 1000, outcome=outcome, error=error, is_final_attempt=is_final_attempt, ) def _publish_embedding_event( *, provider: str, model: str, input_count: int, input_tokens_estimate: int, duration_ms: float, outcome: Literal["success", "error", "cancelled"], error: BaseException | None, is_final_attempt: bool, ) -> None: """Build and emit the EmbeddingCallCompletedEvent. Best-effort.""" try: from src.telemetry.events import ( EmbeddingCallCompletedEvent, EmbeddingCallPurpose, emit, ) from src.utils.types import ( get_embedding_call_purpose, get_embedding_parent_category, get_embedding_run_id, get_embedding_session_id, get_embedding_workspace_name, ) # call_purpose travels via ContextVar so embedding callers don't have # to thread it through every call site. Unknown values drop to None # rather than raising — keeps telemetry resilient to drift. purpose_slug = get_embedding_call_purpose() call_purpose: EmbeddingCallPurpose | None = None if purpose_slug: try: call_purpose = EmbeddingCallPurpose(purpose_slug) except ValueError: logger.debug( "Unknown embedding_call_purpose=%r; emitting without", purpose_slug, ) emit( EmbeddingCallCompletedEvent( workspace_name=get_embedding_workspace_name(), call_purpose=call_purpose, parent_category=get_embedding_parent_category(), provider=provider, model=model, input_count=input_count, input_tokens_estimate=input_tokens_estimate, duration_ms=duration_ms, outcome=outcome, is_final_attempt=is_final_attempt, error_class=type(error).__name__ if error is not None else None, run_id=get_embedding_run_id(), ) ) # Trace stream (ground-truth) — gated on payload tracing. Each embedding # gets its own span nested under the driving agent run (parent_span_id = # run_id), so multiple embeddings in one run don't share a span id. if settings.TELEMETRY.TRACE_PAYLOADS_ENABLED: from src.telemetry.events import EmbeddingCallTracedEvent, emit_trace run_id = get_embedding_run_id() span_id = generate_nanoid() emit_trace( EmbeddingCallTracedEvent( trace_id=run_id or span_id, span_id=span_id, parent_span_id=run_id, session_id=get_embedding_session_id(), call_purpose=purpose_slug, parent_category=get_embedding_parent_category(), provider=provider, model=model, provider_input_tokens=input_tokens_estimate, provider_output_tokens=0, input_count=input_count, ) ) except Exception: # pragma: no cover - telemetry must not raise logger.debug("Failed to emit EmbeddingCallCompletedEvent", exc_info=True) class BatchItem(NamedTuple): """A single item in a batch with its metadata.""" text: str text_id: str chunk_index: int token_count: int class _EmbeddingClient: """ Embedding client supporting OpenAI and Gemini with chunking and batching support. """ def __init__( self, config: EmbeddingModelConfig, *, vector_dimensions: int, max_input_tokens: int, max_tokens_per_request: int, send_dimensions: bool, ): self.transport: str = config.transport self.model: str = config.model self.vector_dimensions: int = vector_dimensions self.send_dimensions: bool = send_dimensions if self.transport == "gemini": if not config.api_key: raise ValueError("Gemini API key is required") # 10-minute HTTP timeout, in lockstep with the LLM registry's Gemini # client (`src/llm/registry.py:_build_gemini_http_options`). Without # this, a stalled Gemini embedding socket wedges the deriver worker # exactly the way #785 describes for the LLM client. http_options = genai_types.HttpOptions( base_url=config.base_url, timeout=600_000, ) self.client: genai.Client | AsyncOpenAI = genai.Client( api_key=config.api_key, http_options=http_options, ) # Gemini has a 2048 token limit self.max_embedding_tokens: int = min(max_input_tokens, 2048) # Gemini batch size is not documented, using conservative estimate self.max_batch_size: int = config.max_batch_size or 100 else: # openai if not config.api_key: raise ValueError("OpenAI API key is required") self.client = AsyncOpenAI( api_key=config.api_key, base_url=config.base_url, ) self.max_embedding_tokens = max_input_tokens self.max_batch_size = config.max_batch_size or 2048 try: self.encoding: tiktoken.Encoding = tiktoken.encoding_for_model(self.model) except KeyError: self.encoding = tiktoken.get_encoding("cl100k_base") self.max_embedding_tokens_per_request: int = max_tokens_per_request @property def provider(self) -> str: return self.transport def _validate_embedding_dimensions(self, embedding: list[float]) -> list[float]: if len(embedding) != self.vector_dimensions: raise ValueError( f"Embedding dimension mismatch for {self.transport}:{self.model}. " + f"Expected {self.vector_dimensions}, got {len(embedding)}." ) return embedding async def embed(self, query: str) -> list[float]: token_count = len(self.encoding.encode(query)) if token_count > self.max_embedding_tokens: raise ValueError( f"Query exceeds maximum token limit of {self.max_embedding_tokens} tokens (got {token_count} tokens)" ) # Bind the typed client at the dispatch site so pyright can narrow it # for the closures without needing `assert isinstance(...)` (bandit # B101). The closures close over the narrowed local, not `self.client`. if isinstance(self.client, genai.Client): gemini_client = self.client async def _call_gemini() -> list[float]: response = await gemini_client.aio.models.embed_content( model=self.model, contents=query, config={"output_dimensionality": self.vector_dimensions}, ) if not response.embeddings or not response.embeddings[0].values: raise ValueError("No embedding returned from Gemini API") return self._validate_embedding_dimensions( response.embeddings[0].values ) return await _emit_embedding_call( provider=self.transport, model=self.model, texts=[query], input_tokens_estimate=token_count, fn=_call_gemini, ) openai_client = self.client async def _call_openai() -> list[float]: openai_kwargs: dict[str, Any] = {"model": self.model, "input": [query]} if self.send_dimensions: openai_kwargs["dimensions"] = self.vector_dimensions response = await openai_client.embeddings.create(**openai_kwargs) return self._validate_embedding_dimensions(response.data[0].embedding) return await _emit_embedding_call( provider=self.transport, model=self.model, texts=[query], input_tokens_estimate=token_count, fn=_call_openai, ) async def simple_batch_embed(self, texts: list[str]) -> list[list[float]]: """ Batch-embed a list of text strings. Each input must already fit within `max_embedding_tokens`; this method does not sub-chunk oversized inputs. Internally goes through the same token-aware batching pipeline as `batch_embed()` so the per-request token cap is respected. Args: texts: List of text strings to embed Returns: List of embedding vectors, one per input text (in order) Raises: ValueError: If any text exceeds token limits """ if not texts: return [] # Validate per-input token limit and collect token counts for batching token_counts: list[int] = [] for idx, text in enumerate(texts): tokens = len(self.encoding.encode(text)) if tokens > self.max_embedding_tokens: raise ValueError( f"Text at index {idx} exceeds maximum token limit of {self.max_embedding_tokens} tokens (got {tokens} tokens)" ) token_counts.append(tokens) # Use positional indices as text_ids so we can reassemble in input order. text_chunks: dict[str, list[tuple[str, int]]] = { str(i): [(text, token_counts[i])] for i, text in enumerate(texts) } batches = self._create_batches(text_chunks) batch_results = await asyncio.gather( *[self._process_batch(batch) for batch in batches], ) combined: dict[str, list[list[float]]] = self._accumulate_embeddings( batch_results ) return [combined[str(i)][0] for i in range(len(texts))] def prepare_chunks(self, id_resource_dict: dict[str, str]) -> dict[str, list[str]]: """ Public helper: tokenize and chunk texts using the same rules as `batch_embed()`. Returns ordered chunk texts per input id. Intended for callers that want to persist embeddable chunks before later embedding them off the request path. """ return { text_id: [chunk_text for chunk_text, _ in chunks] for text_id, chunks in self._prepare_chunks(id_resource_dict).items() } async def batch_embed( self, id_resource_dict: dict[str, str] ) -> dict[str, list[list[float]]]: """ Embed multiple texts, chunking long ones and batching API calls. Args: id_resource_dict: Maps text IDs to text content Returns: Maps text IDs to lists of embedding vectors (one per chunk) """ if not id_resource_dict: return {} # 1. Prepare chunks for all texts if needed text_chunks = self._prepare_chunks(id_resource_dict) # 2. Create batches that fit API limits (max 2048 embeddings per request, max 300,000 tokens per request) batches = self._create_batches(text_chunks) # 3. Process all batches concurrently batch_results = await asyncio.gather( *[self._process_batch(batch) for batch in batches], ) # 4. Accumulate results preserving chunk order return self._accumulate_embeddings(batch_results) def _prepare_chunks( self, id_resource_dict: dict[str, str] ) -> dict[str, list[tuple[str, int]]]: """ Chunk texts that exceed token limits. Args: id_resource_dict: Maps text IDs to text content. We tokenize with the embedding client's own encoding so token IDs match the decoder vocabulary used by the target embedding API. Returns: Maps text IDs to lists of (chunk_text, token_count) tuples """ out: dict[str, list[tuple[str, int]]] = {} for text_id, text in id_resource_dict.items(): tokens = self.encoding.encode(text) if len(tokens) > self.max_embedding_tokens: out[text_id] = _chunk_text_with_tokens( text, tokens, self.max_embedding_tokens, self.encoding ) else: out[text_id] = [(text, len(tokens))] return out def _create_batches( self, text_chunks: dict[str, list[tuple[str, int]]] ) -> list[list[BatchItem]]: """ Group chunks into batches that fit API limits. Args: text_chunks: Maps text IDs to lists of (chunk_text, token_count) tuples Returns: List of batches, each containing BatchItem objects """ batches: list[list[BatchItem]] = [] current_batch: list[BatchItem] = [] current_tokens = 0 for text_id, chunks in text_chunks.items(): for chunk_idx, (chunk_text, chunk_tokens) in enumerate(chunks): # Check if adding this chunk would exceed limits would_exceed_tokens = ( current_tokens + chunk_tokens > self.max_embedding_tokens_per_request ) would_exceed_count = len(current_batch) >= self.max_batch_size if current_batch and (would_exceed_tokens or would_exceed_count): batches.append(current_batch) current_batch = [] current_tokens = 0 current_batch.append( BatchItem(chunk_text, text_id, chunk_idx, chunk_tokens) ) current_tokens += chunk_tokens if current_batch: batches.append(current_batch) return batches async def _process_batch( self, batch: list[BatchItem], max_retries: int = 3 ) -> dict[str, dict[int, list[float]]]: """ Process a single batch through the embeddings API with retry logic. Args: batch: List of BatchItem objects to embed max_retries: Maximum number of retry attempts (default: 3) Returns: Maps text IDs to {chunk_index: embedding_vector} dictionaries """ last_exception: Exception | None = None async def _call_provider() -> dict[str, dict[int, list[float]]]: """One provider call. Lifted out of the retry loop so _emit_embedding_call emits a separate event per attempt — each attempt is a distinct provider hit and shows up as its own line item in analytics.""" result: dict[str, dict[int, list[float]]] = defaultdict(dict) if isinstance(self.client, genai.Client): response = await self.client.aio.models.embed_content( model=self.model, contents=[item.text for item in batch], config={"output_dimensionality": self.vector_dimensions}, ) if response.embeddings: for item, embedding in zip(batch, response.embeddings, strict=True): if embedding.values: result[item.text_id][item.chunk_index] = ( self._validate_embedding_dimensions(embedding.values) ) else: # openai openai_kwargs: dict[str, Any] = { "model": self.model, "input": [item.text for item in batch], } if self.send_dimensions: openai_kwargs["dimensions"] = self.vector_dimensions response = await self.client.embeddings.create(**openai_kwargs) for item, embedding_data in zip(batch, response.data, strict=True): result[item.text_id][item.chunk_index] = ( self._validate_embedding_dimensions(embedding_data.embedding) ) return result # Token counts were computed during chunk prep; reuse them here so the # provider call doesn't re-encode every chunk just for the size proxy. batch_tokens_estimate = sum(item.token_count for item in batch) batch_texts = [item.text for item in batch] for attempt in range(max_retries): try: result = await _emit_embedding_call( provider=self.transport, model=self.model, texts=batch_texts, input_tokens_estimate=batch_tokens_estimate, fn=_call_provider, is_final_attempt=(attempt >= max_retries - 1), ) return dict(result) except Exception as e: last_exception = e if attempt < max_retries - 1: # Exponential backoff: 1s, 2s, 4s wait_time = 2**attempt logger.warning( f"Embedding batch failed (attempt {attempt + 1}/{max_retries}), " + f"retrying in {wait_time}s: {e}" ) await asyncio.sleep(wait_time) else: logger.exception("Error processing batch after all retries") raise last_exception or RuntimeError("Batch processing failed") def _accumulate_embeddings( self, batch_results: list[dict[str, dict[int, list[float]]]] ) -> dict[str, list[list[float]]]: """ Combine batch results into final output, preserving chunk order. Args: batch_results: List of batch results from _process_batch Returns: Maps text IDs to ordered lists of embedding vectors """ all_embeddings: dict[str, dict[int, list[float]]] = defaultdict(dict) # Collect all embeddings by text_id and chunk_index for batch_result in batch_results: for text_id, chunk_dict in batch_result.items(): all_embeddings[text_id].update(chunk_dict) # Convert to ordered lists return { text_id: [chunk_dict[i] for i in sorted(chunk_dict.keys())] for text_id, chunk_dict in all_embeddings.items() } def _chunk_text_with_tokens( text: str, encoded_tokens: list[int], max_tokens: int, encoding: tiktoken.Encoding, ) -> list[tuple[str, int]]: """ Split text into chunks that fit within token limits, with 20% overlap. Args: text: Original text to chunk encoded_tokens: Pre-encoded tokens for the text max_tokens: Maximum tokens per chunk encoding: Tiktoken encoding model Returns: List of (chunk_text, token_count) tuples """ if len(encoded_tokens) <= max_tokens: return [(text, len(encoded_tokens))] # Use 20% overlap for better semantic continuity overlap_tokens = int(max_tokens * 0.2) step_size = max_tokens - overlap_tokens return [ ( encoding.decode(encoded_tokens[i : i + max_tokens]), min(max_tokens, len(encoded_tokens) - i), ) for i in range(0, len(encoded_tokens), step_size) if i < len(encoded_tokens) # Ensure we don't create empty chunks ] class EmbeddingClient: """ Singleton wrapper for the embedding client with deferred loading. The actual client is only initialized on first use, improving startup time and allowing the application to start even if API keys are not yet configured. """ _instance: "_EmbeddingClient | None" = None _instance_signature: tuple[object, ...] | None = None _lock: threading.Lock = threading.Lock() _wrapper_instance: "EmbeddingClient | None" = None def __new__(cls): """Ensure only one instance of EmbeddingClient exists.""" # We always return the same wrapper instance if cls._wrapper_instance is None: cls._wrapper_instance = super().__new__(cls) return cls._wrapper_instance def _get_client(self) -> _EmbeddingClient: """ Get or create the underlying embedding client instance. Uses double-checked locking for thread-safe lazy initialization. """ signature = self._get_settings_signature() if self._instance is None or self._instance_signature != signature: with self._lock: if self._instance is None or self._instance_signature != signature: runtime_config = self._resolve_runtime_config() self._instance = _EmbeddingClient( runtime_config, vector_dimensions=settings.EMBEDDING.VECTOR_DIMENSIONS, max_input_tokens=settings.EMBEDDING.MAX_INPUT_TOKENS, max_tokens_per_request=settings.EMBEDDING.MAX_TOKENS_PER_REQUEST, send_dimensions=settings.EMBEDDING.resolve_send_dimensions(), ) self._instance_signature = signature logger.debug( "Initialized embedding client with transport: %s model: %s", runtime_config.transport, runtime_config.model, ) return self._instance def _resolve_runtime_config(self) -> EmbeddingModelConfig: return resolve_embedding_model_config(settings.EMBEDDING.MODEL_CONFIG) def _get_settings_signature(self) -> tuple[object, ...]: runtime_config = self._resolve_runtime_config() return ( runtime_config.transport, runtime_config.model, runtime_config.api_key, runtime_config.base_url, runtime_config.max_batch_size, settings.EMBEDDING.VECTOR_DIMENSIONS, settings.EMBEDDING.MAX_INPUT_TOKENS, settings.EMBEDDING.MAX_TOKENS_PER_REQUEST, settings.EMBEDDING.resolve_send_dimensions(), ) async def embed(self, query: str) -> list[float]: """Embed a single query string.""" return await self._get_client().embed(query) async def simple_batch_embed(self, texts: list[str]) -> list[list[float]]: """Batch embed a list of text strings (each must fit token limit).""" return await self._get_client().simple_batch_embed(texts) def prepare_chunks(self, id_resource_dict: dict[str, str]) -> dict[str, list[str]]: """Chunk texts using the same rules as `batch_embed` (no network).""" return self._get_client().prepare_chunks(id_resource_dict) async def batch_embed( self, id_resource_dict: dict[str, str] ) -> dict[str, list[list[float]]]: """Embed multiple texts, chunking long ones and batching API calls.""" return await self._get_client().batch_embed(id_resource_dict) @property def provider(self) -> str: """Get the provider name.""" return self._get_client().provider @property def model(self) -> str: """Get the model name.""" return self._get_client().model @property def transport(self) -> str: """Get the transport name.""" return self._get_client().transport @property def max_embedding_tokens(self) -> int: """Get the maximum embedding tokens.""" return self._get_client().max_embedding_tokens @property def vector_dimensions(self) -> int: """Get the configured embedding dimensions.""" return self._get_client().vector_dimensions @property def encoding(self) -> tiktoken.Encoding: """Get the tiktoken encoding. Resolved without constructing the underlying client: tiktoken needs no API key, and token-counting callers (e.g. the document dedup tie-break) must work in environments with no embedding credentials, such as CI for pull requests from forks. """ if self._instance is not None: return self._instance.encoding try: return tiktoken.encoding_for_model(self._resolve_runtime_config().model) except KeyError: return tiktoken.get_encoding("cl100k_base") # Shared singleton embedding client instance embedding_client = EmbeddingClient()