Add response token count logic to Bedrock instrumentation. #1504
Conversation
umaannamalai
changed the title
✅MegaLinter analysis: Success
See detailed reports in MegaLinter artifacts
|
Codecov Report❌ Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## feature-llm-token-counts #1504 +/- ##
===========================================================
Coverage ? 81.65%
===========================================================
Files ? 207
Lines ? 24037
Branches ? 3802
===========================================================
Hits ? 19628
Misses ? 3132
Partials ? 1277 ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
There was a problem hiding this comment.
I said this in a comment too but I'm going to stop reviewing this for the moment.
Can you for the sake of being able to easily identify what's changed here: re-order these changes in the file such that they end up being diffed against the original function. Right now it's really hard to compare to see what's changed because the diff isn't showing a 1:1 comparison. I know it's a pain but in a larger PR like this with lots of files changed it makes it a ton easier to review.
|
|
||
| bedrock_attrs["input"] = request_body.get("inputText") | ||
|
|
||
| def extract_bedrock_titan_text_model_streaming_response(response_body, bedrock_attrs): |
There was a problem hiding this comment.
Can you for the sake of being able to easily identify what's changed here: re-order these changes in the file such that they end up being diffed against the original function. Right now it's really hard to compare to see what's changed because the diff isn't showing a 1:1 comparison. I know it's a pain but in a larger PR like this with lots of files changed it makes it a ton easier to review.
| model = bedrock_attrs.get("model", None) | ||
| input_ = bedrock_attrs.get("input") | ||
|
|
||
| response_total_tokens = bedrock_attrs.get("response.usage.total_tokens", None) |
There was a problem hiding this comment.
None is redundant.
| response_total_tokens = bedrock_attrs.get("response.usage.total_tokens", None) | |
| response_total_tokens = bedrock_attrs.get("response.usage.total_tokens") |
| response_id = bedrock_attrs.get("response_id", None) | ||
| model = bedrock_attrs.get("model", None) | ||
|
|
||
| response_prompt_tokens = bedrock_attrs.get("response.usage.prompt_tokens", None) |
There was a problem hiding this comment.
| response_prompt_tokens = bedrock_attrs.get("response.usage.prompt_tokens", None) | |
| response_prompt_tokens = bedrock_attrs.get("response.usage.prompt_tokens") |
| model = bedrock_attrs.get("model", None) | ||
|
|
||
| response_prompt_tokens = bedrock_attrs.get("response.usage.prompt_tokens", None) | ||
| response_completion_tokens = bedrock_attrs.get("response.usage.completion_tokens", None) |
There was a problem hiding this comment.
| response_completion_tokens = bedrock_attrs.get("response.usage.completion_tokens", None) | |
| response_completion_tokens = bedrock_attrs.get("response.usage.completion_tokens") |
|
|
||
| response_prompt_tokens = bedrock_attrs.get("response.usage.prompt_tokens", None) | ||
| response_completion_tokens = bedrock_attrs.get("response.usage.completion_tokens", None) | ||
| response_total_tokens = bedrock_attrs.get("response.usage.total_tokens", None) |
There was a problem hiding this comment.
| response_total_tokens = bedrock_attrs.get("response.usage.total_tokens", None) | |
| response_total_tokens = bedrock_attrs.get("response.usage.total_tokens") |
| len(input_message_list) + len(output_message_list) | ||
| ) or None # If 0, attribute will be set to None and removed | ||
|
|
||
| input_message_content = " ".join([msg.get("content", "") for msg in input_message_list if msg.get("content")]) |
There was a problem hiding this comment.
I don't think the default values of "" are reachable here given the if msg.get("content").
| "ingest_source": "Python", | ||
| } | ||
| if all_token_counts: | ||
| chat_completion_message_dict["token_count"] = 0 |
There was a problem hiding this comment.
Is this the value that was previously added in the pipeline? Are we just setting it to 0 as a transition?
This PR updates sync and async Bedrock instrumentation and tests to pull token counts directly from the response object. If a customer has an LLM token counting callback function registered, this will still take priority over the response token count values.
It also removes token count calculations from error cases as the agent will consistently stop reporting token counts when a error was met during the creation of a text generation.