feat(ai-monitoring): Cohere integration

.github/workflows/test-integrations-data-processing.yml

@@ -58,6 +58,10 @@ jobs:
         run: |
           set -x # print commands that are executed
           ./scripts/runtox.sh "py${{ matrix.python-version }}-celery-latest" --cov=tests --cov=sentry_sdk --cov-report= --cov-branch
+      - name: Test cohere latest
+        run: |
+          set -x # print commands that are executed
+          ./scripts/runtox.sh "py${{ matrix.python-version }}-cohere-latest" --cov=tests --cov=sentry_sdk --cov-report= --cov-branch
       - name: Test huey latest
         run: |
           set -x # print commands that are executed
@@ -126,6 +130,10 @@ jobs:
         run: |
           set -x # print commands that are executed
           ./scripts/runtox.sh --exclude-latest "py${{ matrix.python-version }}-celery" --cov=tests --cov=sentry_sdk --cov-report= --cov-branch
+      - name: Test cohere pinned
+        run: |
+          set -x # print commands that are executed
+          ./scripts/runtox.sh --exclude-latest "py${{ matrix.python-version }}-cohere" --cov=tests --cov=sentry_sdk --cov-report= --cov-branch
       - name: Test huey pinned
         run: |
           set -x # print commands that are executed

mypy.ini

@@ -25,7 +25,8 @@ warn_unused_ignores = True
 ;
 ; Do not use wildcards in module paths, otherwise added modules will
 ; automatically have the same set of relaxed rules as the rest
-
+[mypy-cohere.*]
+ignore_missing_imports = True
 [mypy-django.*]
 ignore_missing_imports = True
 [mypy-pyramid.*]

scripts/split-tox-gh-actions/split-tox-gh-actions.py

@@ -70,6 +70,7 @@
         "arq",
         "beam",
         "celery",
+        "cohere",
         "huey",
         "langchain",
         "openai",

sentry_sdk/consts.py

@@ -91,6 +91,18 @@ class SPANDATA:
     See: https://develop.sentry.dev/sdk/performance/span-data-conventions/
     """
 
+    AI_FREQUENCY_PENALTY = "ai.frequency_penalty"
+    """
+    Used to reduce repetitiveness of generated tokens.
+    Example: 0.5
+    """
+
+    AI_PRESENCE_PENALTY = "ai.presence_penalty"
+    """
+    Used to reduce repetitiveness of generated tokens.
+    Example: 0.5
+    """
+
     AI_INPUT_MESSAGES = "ai.input_messages"
     """
     The input messages to an LLM call.
@@ -164,12 +176,31 @@ class SPANDATA:
     For an AI model call, the logit bias
     """
 
+    AI_PREAMBLE = "ai.preamble"
+    """
+    For an AI model call, the preamble parameter.
+    Preambles are a part of the prompt used to adjust the model's overall behavior and conversation style.
+    Example: "You are now a clown."
+    """
+
+    AI_RAW_PROMPTING = "ai.raw_prompting"
+    """
+    Minimize pre-processing done to the prompt sent to the LLM.
+    Example: true
+    """
+
     AI_RESPONSES = "ai.responses"
     """
     The responses to an AI model call. Always as a list.
     Example: ["hello", "world"]
     """
 
+    AI_SEED = "ai.seed"
+    """
+    The seed, ideally models given the same seed and same other parameters will produce the exact same output.
+    Example: 123.45
+    """
+
     DB_NAME = "db.name"
     """
     The name of the database being accessed. For commands that switch the database, this should be set to the target database (even if the command fails).
@@ -298,6 +329,8 @@ class SPANDATA:
 class OP:
     ANTHROPIC_MESSAGES_CREATE = "ai.messages.create.anthropic"
     CACHE_GET_ITEM = "cache.get_item"
+    COHERE_CHAT_COMPLETIONS_CREATE = "ai.chat_completions.create.cohere"
+    COHERE_EMBEDDINGS_CREATE = "ai.embeddings.create.cohere"
     DB = "db"
     DB_REDIS = "db.redis"
     EVENT_DJANGO = "event.django"

sentry_sdk/integrations/__init__.py

@@ -78,6 +78,7 @@ def iter_default_integrations(with_auto_enabling_integrations):
     "sentry_sdk.integrations.celery.CeleryIntegration",
     "sentry_sdk.integrations.chalice.ChaliceIntegration",
     "sentry_sdk.integrations.clickhouse_driver.ClickhouseDriverIntegration",
+    "sentry_sdk.integrations.cohere.CohereIntegration",
     "sentry_sdk.integrations.django.DjangoIntegration",
     "sentry_sdk.integrations.falcon.FalconIntegration",
     "sentry_sdk.integrations.fastapi.FastApiIntegration",

sentry_sdk/integrations/cohere.py

@@ -0,0 +1,257 @@
+from functools import wraps
+
+from sentry_sdk import consts
+from sentry_sdk._types import TYPE_CHECKING
+from sentry_sdk.ai.monitoring import record_token_usage
+from sentry_sdk.consts import SPANDATA
+from sentry_sdk.ai.utils import set_data_normalized
+
+if TYPE_CHECKING:
+    from typing import Any, Callable, Iterator
+    from sentry_sdk.tracing import Span
+
+import sentry_sdk
+from sentry_sdk.scope import should_send_default_pii
+from sentry_sdk.integrations import DidNotEnable, Integration
+from sentry_sdk.utils import (
+    capture_internal_exceptions,
+    event_from_exception,
+    ensure_integration_enabled,
+)
+
+try:
+    from cohere.client import Client
+    from cohere.base_client import BaseCohere
+    from cohere import ChatStreamEndEvent, NonStreamedChatResponse
+
+    if TYPE_CHECKING:
+        from cohere import StreamedChatResponse
+except ImportError:
+    raise DidNotEnable("Cohere not installed")
+
+
+COLLECTED_CHAT_PARAMS = {
+    "model": SPANDATA.AI_MODEL_ID,
+    "k": SPANDATA.AI_TOP_K,
+    "p": SPANDATA.AI_TOP_P,
+    "seed": SPANDATA.AI_SEED,
+    "frequency_penalty": SPANDATA.AI_FREQUENCY_PENALTY,
+    "presence_penalty": SPANDATA.AI_PRESENCE_PENALTY,
+    "raw_prompting": SPANDATA.AI_RAW_PROMPTING,
+}
+
+COLLECTED_PII_CHAT_PARAMS = {
+    "tools": SPANDATA.AI_TOOLS,
+    "preamble": SPANDATA.AI_PREAMBLE,
+}
+
+COLLECTED_CHAT_RESP_ATTRS = {
+    "generation_id": "ai.generation_id",
+    "is_search_required": "ai.is_search_required",
+    "finish_reason": "ai.finish_reason",
+}
+
+COLLECTED_PII_CHAT_RESP_ATTRS = {
+    "citations": "ai.citations",
+    "documents": "ai.documents",
+    "search_queries": "ai.search_queries",
+    "search_results": "ai.search_results",
+    "tool_calls": "ai.tool_calls",
+}
+
+
+class CohereIntegration(Integration):
+    identifier = "cohere"
+
+    def __init__(self, include_prompts=True):
+        # type: (CohereIntegration, bool) -> None
+        self.include_prompts = include_prompts
+
+    @staticmethod
+    def setup_once():
+        # type: () -> None
+        BaseCohere.chat = _wrap_chat(BaseCohere.chat, streaming=False)
+        Client.embed = _wrap_embed(Client.embed)
+        BaseCohere.chat_stream = _wrap_chat(BaseCohere.chat_stream, streaming=True)
+
+
+def _capture_exception(exc):
+    # type: (Any) -> None
+    event, hint = event_from_exception(
+        exc,
+        client_options=sentry_sdk.get_client().options,
+        mechanism={"type": "cohere", "handled": False},
+    )
+    sentry_sdk.capture_event(event, hint=hint)
+
+
+def _wrap_chat(f, streaming):
+    # type: (Callable[..., Any], bool) -> Callable[..., Any]
+
+    def collect_chat_response_fields(span, res, include_pii):
+        # type: (Span, NonStreamedChatResponse, bool) -> None
+        if include_pii:
+            if hasattr(res, "text"):
+                set_data_normalized(
+                    span,
+                    SPANDATA.AI_RESPONSES,
+                    [res.text],
+                )
+            for pii_attr in COLLECTED_PII_CHAT_RESP_ATTRS:
+                if hasattr(res, pii_attr):
+                    set_data_normalized(span, "ai." + pii_attr, getattr(res, pii_attr))
+
+        for attr in COLLECTED_CHAT_RESP_ATTRS:
+            if hasattr(res, attr):
+                set_data_normalized(span, "ai." + attr, getattr(res, attr))
+
+        if hasattr(res, "meta"):
+            if hasattr(res.meta, "billed_units"):
+                record_token_usage(
+                    span,
+                    prompt_tokens=res.meta.billed_units.input_tokens,
+                    completion_tokens=res.meta.billed_units.output_tokens,
+                )
+            elif hasattr(res.meta, "tokens"):
+                record_token_usage(
+                    span,
+                    prompt_tokens=res.meta.tokens.input_tokens,
+                    completion_tokens=res.meta.tokens.output_tokens,
+                )
+
+            if hasattr(res.meta, "warnings"):
+                set_data_normalized(span, "ai.warnings", res.meta.warnings)
+
+    @wraps(f)
+    @ensure_integration_enabled(CohereIntegration, f)
+    def new_chat(*args, **kwargs):
+        # type: (*Any, **Any) -> Any
+        if "message" not in kwargs:
+            return f(*args, **kwargs)
+
+        if not isinstance(kwargs.get("message"), str):
+            return f(*args, **kwargs)
+
+        message = kwargs.get("message")
+
+        span = sentry_sdk.start_span(
+            op=consts.OP.COHERE_CHAT_COMPLETIONS_CREATE,
+            description="cohere.client.Chat",
+        )
+        span.__enter__()
+        try:
+            res = f(*args, **kwargs)
+        except Exception as e:
+            _capture_exception(e)
+            span.__exit__(None, None, None)
+            raise e from None
+
+        integration = sentry_sdk.get_client().get_integration(CohereIntegration)
+
+        with capture_internal_exceptions():
+            if should_send_default_pii() and integration.include_prompts:
+                set_data_normalized(
+                    span,
+                    SPANDATA.AI_INPUT_MESSAGES,
+                    list(
+                        map(
+                            lambda x: {
+                                "role": getattr(x, "role", "").lower(),
+                                "content": getattr(x, "message", ""),
+                            },
+                            kwargs.get("chat_history", []),
+                        )
+                    )
+                    + [{"role": "user", "content": message}],
+                )
+                for k, v in COLLECTED_PII_CHAT_PARAMS.items():
+                    if k in kwargs:
+                        set_data_normalized(span, v, kwargs[k])
+
+            for k, v in COLLECTED_CHAT_PARAMS.items():
+                if k in kwargs:
+                    set_data_normalized(span, v, kwargs[k])
+            set_data_normalized(span, SPANDATA.AI_STREAMING, False)
+
+            if streaming:
+                old_iterator = res
+
+                def new_iterator():
+                    # type: () -> Iterator[StreamedChatResponse]
+
+                    with capture_internal_exceptions():
+                        for x in old_iterator:
+                            if isinstance(x, ChatStreamEndEvent):
+                                collect_chat_response_fields(
+                                    span,
+                                    x.response,
+                                    include_pii=should_send_default_pii()
+                                    and integration.include_prompts,
+                                )
+                            yield x
+
+                    span.__exit__(None, None, None)
+
+                return new_iterator()
+            elif isinstance(res, NonStreamedChatResponse):
+                collect_chat_response_fields(
+                    span,
+                    res,
+                    include_pii=should_send_default_pii()
+                    and integration.include_prompts,
+                )
+                span.__exit__(None, None, None)
+            else:
+                set_data_normalized(span, "unknown_response", True)
+                span.__exit__(None, None, None)
+            return res
+
+    return new_chat
+
+
+def _wrap_embed(f):
+    # type: (Callable[..., Any]) -> Callable[..., Any]
+
+    @wraps(f)
+    @ensure_integration_enabled(CohereIntegration, f)
+    def new_embed(*args, **kwargs):
+        # type: (*Any, **Any) -> Any
+        with sentry_sdk.start_span(
+            op=consts.OP.COHERE_EMBEDDINGS_CREATE,
+            description="Cohere Embedding Creation",
+        ) as span:
+            integration = sentry_sdk.get_client().get_integration(CohereIntegration)
+            if "texts" in kwargs and (
+                should_send_default_pii() and integration.include_prompts
+            ):
+                if isinstance(kwargs["texts"], str):
+                    set_data_normalized(span, "ai.texts", [kwargs["texts"]])
+                elif (
+                    isinstance(kwargs["texts"], list)
+                    and len(kwargs["texts"]) > 0
+                    and isinstance(kwargs["texts"][0], str)
+                ):
+                    set_data_normalized(
+                        span, SPANDATA.AI_INPUT_MESSAGES, kwargs["texts"]
+                    )
+
+            if "model" in kwargs:
+                set_data_normalized(span, SPANDATA.AI_MODEL_ID, kwargs["model"])
+            try:
+                res = f(*args, **kwargs)
+            except Exception as e:
+                _capture_exception(e)
+                raise e from None
+            if (
+                hasattr(res, "meta")
+                and hasattr(res.meta, "billed_units")
+                and hasattr(res.meta.billed_units, "input_tokens")
+            ):
+                record_token_usage(
+                    span,
+                    prompt_tokens=res.meta.billed_units.input_tokens,
+                    total_tokens=res.meta.billed_units.input_tokens,
+                )
+            return res
+
+    return new_embed

sentry_sdk/integrations/langchain.py

@@ -63,7 +63,12 @@ def count_tokens(s):
 
 # To avoid double collecting tokens, we do *not* measure
 # token counts for models for which we have an explicit integration
-NO_COLLECT_TOKEN_MODELS = ["openai-chat"]  # TODO add huggingface and anthropic
+NO_COLLECT_TOKEN_MODELS = [
+    "openai-chat",
+    "anthropic-chat",
+    "cohere-chat",
+    "huggingface_endpoint",
+]
 
 
 class LangchainIntegration(Integration):
@@ -216,6 +221,7 @@ def on_chat_model_start(self, serialized, messages, *, run_id, **kwargs):
             watched_span.no_collect_tokens = any(
                 x in all_params.get("_type", "") for x in NO_COLLECT_TOKEN_MODELS
             )
+
             if not model and "anthropic" in all_params.get("_type"):
                 model = "claude-2"
             if model:

tests/integrations/cohere/__init__.py

@@ -0,0 +1,3 @@
+import pytest
+
+pytest.importorskip("cohere")