Add basic Open Telemetry instrumentation for model calls.

This commit wraps all LLM model calls in an Open Telemetry span that
abides by the (still nascent) semantic conventions for Generative AI
clients [0].

It's very similar in approach to what was done for `httr2`, and in fact
the two of them complement one another nicely:
r-lib/httr2#729.

For example:

    library(otelsdk)

    Sys.setenv(OTEL_TRACES_EXPORTER = "stderr")

    chat <- ellmer::chat_databricks(model = "databricks-claude-3-7-sonnet")
    chat$chat("Tell me a joke in the form of an SQL query.")

[0]: https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-spans/

Signed-off-by: Aaron Jacobs <[email protected]>

Author: atheriel

DESCRIPTION

@@ -41,6 +41,7 @@ Suggests:
     knitr,
     magick,
     openssl,
+    otel (>= 0.0.0.9000),
     paws.common,
     rmarkdown,
     shiny,
@@ -109,3 +110,5 @@ Collate:
     'utils-prettytime.R'
     'utils.R'
     'zzz.R'
+Remotes:  
+    r-lib/otel

R/chat.R

@@ -627,7 +627,7 @@ Chat <- R6::R6Class(
       if (echo == "all") {
         cat_line(format(user_turn), prefix = "> ")
       }
-
+      span <- start_chat_span(private$provider)
       response <- chat_perform(
         provider = private$provider,
         mode = if (stream) "stream" else "value",
@@ -654,9 +654,11 @@ Chat <- R6::R6Class(
 
           result <- stream_merge_chunks(private$provider, result, chunk)
         }
+        end_chat_span(span, result)
         turn <- value_turn(private$provider, result, has_type = !is.null(type))
         turn <- match_tools(turn, private$tools)
       } else {
+        end_chat_span(span, response)
         turn <- value_turn(
           private$provider,
           response,
@@ -709,6 +711,7 @@ Chat <- R6::R6Class(
       type = NULL,
       yield_as_content = FALSE
     ) {
+      span <- start_chat_span(private$provider)
       response <- chat_perform(
         provider = private$provider,
         mode = if (stream) "async-stream" else "async-value",
@@ -735,10 +738,12 @@ Chat <- R6::R6Class(
 
           result <- stream_merge_chunks(private$provider, result, chunk)
         }
+        end_chat_span(span, result)
         turn <- value_turn(private$provider, result, has_type = !is.null(type))
       } else {
         result <- await(response)
 
+        end_chat_span(span, result)
         turn <- value_turn(private$provider, result, has_type = !is.null(type))
         text <- turn@text
         if (!is.null(text)) {

R/otel.R

@@ -0,0 +1,48 @@
+# Starts an Open Telemetry span that abides by the semantic conventions for
+# Generative AI clients.
+#
+# See: https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-spans/
+start_chat_span <- function(provider, tracer = NULL, scope = parent.frame()) {
+  if (!is_installed("otel")) {
+    return(NULL)
+  }
+  tracer <- tracer %||% otel::get_tracer("ellmer")
+  name <- sprintf("chat %s", provider@model)
+  if (!tracer$is_enabled()) {
+    # Return a no-op span when tracing is disabled.
+    return(tracer$start_span(name))
+  }
+  tracer$start_span(
+    name,
+    options = list(kind = "CLIENT"),
+    # Ensure we set attributes relevant to sampling at span creation time.
+    attributes = compact(list(
+      "gen_ai.operation.name" = "chat",
+      "gen_ai.system" = tolower(provider@name),
+      "gen_ai.request.model" = provider@model
+    )),
+    scope = scope
+  )
+}
+
+end_chat_span <- function(span, result) {
+  if (is.null(span) || !span$is_recording()) {
+    return(invisible(span))
+  }
+  if (!is.null(result$model)) {
+    span$set_attribute("gen_ai.response.model", result$model)
+  }
+  if (!is.null(result$id)) {
+    span$set_attribute("gen_ai.response.id", result$id)
+  }
+  if (!is.null(result$usage)) {
+    span$set_attribute("gen_ai.usage.input_tokens", result$usage$prompt_tokens)
+    span$set_attribute(
+      "gen_ai.usage.output_tokens",
+      result$usage$completion_tokens
+    )
+  }
+  # TODO: Consider setting gen_ai.response.finish_reasons.
+  span$set_status("ok")
+  span$end()
+}