[ENHANCEMENT]: Session level attribute filtering with graphql

[saadshaikh3]

Is your feature request related to a problem? Please describe.

It's very problematic for me that I am recording the session attributes in metadata but when retrieving data I am not able to filter on it. So basically my sessions have metadata dictionary with field "type": "AssistantA". I want to retrieve sessions from a project filtered based on attribute.metadata but thats not possible so I have to manage the filtering in backned on my side and then that causes issues with the pagination because sessions include all the records for other assistants as well. So I have to handle pagination in backend too for my dashboards.

Also this forces me to retrieve records in 1000s. Like I have to pass sessions.first: 2000 so that I receive adequate amount of recent sessions from which i can filter the adequate amount of assistants accordingly at backend but this casues issues sometimes. The following error is returned then:

"message": "QueuePool limit of size 5 overflow 10 reached, connection timed out, timeout 30.00 (Background on this error at: https://sqlalche.me/e/20/3o7r)"%22),

Describe the solution you'd like

I see there are options of filters in other objects but for projects.sessions there should be an option to filter from graphql like project.session.attributes.metadata["myfield"]="myvalue" and then the results is return paginated as per this filter.

Having this filter with sort, first and timeRange will make life easy.

Describe any alternative solutions you've considered

If you can get the queuepool limit error fixed then maybe i can atleast pursue the current approach I am at.

Additional context

No response

Social Media Handle

No response

[RogerHYang]

Thanks for filing this issue. Is search bar (as shown below) giving you the spans that you need?

[saadshaikh3]

Thanks @RogerHYang for taking a look at this.
I do see the filtered spans from here on UI. But when I try to use graphql and return the records with filter as well as with rootSpan bool = True, the records still include the other internal spans as well and not just the root spans. Its as if the filter condition overwrites the rootSpansOnly bool. This is my query just for reference to retrieve root spans with sessions for a given time range with metadata filter.

I only kept the input { value } in spans.node to verify that whether this span is a root span or not. I do not wish to keep it in my final query as I retrieve the input from session.

The output of this query is 10 spans which includes internal spans and a root span

query GetSpansWithSessions($projectName: String!, $firstSessions: Int, $after: String, $sessionStart: DateTime, $sessionEnd: DateTime, $sortCol: SpanColumn! = startTime, $sortDir: SortDir! = desc) {
  projects(filter: {value: $projectName, col: name}) {
    edges {
      node {
        spans(
          rootSpansOnly: true
          first: $firstSessions
          sort: {dir: $sortDir, col: $sortCol}
          timeRange: {end: $sessionEnd, start: $sessionStart}
          after: $after
          filterCondition: "metadata['type']=='chat'"
        ) {
          pageInfo {
            endCursor
            hasNextPage
          }
          edges {
            node {
              input {
                value
              }
              metadata
              trace {
                session {
                  firstInput {
                    value
                  }
                  numTraces
                  numTracesWithError
                  costSummary {
                    total {
                      cost
                    }
                  }
                  tokenUsage {
                    total
                  }
                  endTime
                  sessionId
                  startTime
                  traces {
                    edges {
                      node {
                        rootSpan {
                          cumulativeTokenCountTotal
                          id
                          endTime
                          startTime
                          input {
                            value
                          }
                          output {
                            value
                          }
                          latencyMs
                        }
                      }
                    }
                  }
                }
              }
            }
          }
        }
      }
    }
  }
}

[RogerHYang]

I haven't been able to reproduce this issue.

Do you use a different value for $sortCol that isn't startTime?

What do you see if you add orphanSpanAsRootSpan: false as follows?

        spans(
          orphanSpanAsRootSpan: false
          rootSpansOnly: true
          first: $firstSessions
          sort: {dir: $sortDir, col: $sortCol}
          timeRange: {end: $sessionEnd, start: $sessionStart}
          after: $after
          filterCondition: "metadata['type']=='chat'"
        )

What do you see if you add parentId under the span node? Do they show up as null?

          edges {
            node {
              parentId
              input {
                value
              }

[saadshaikh3]

Thanks @RogerHYang. It seems to work now and returns only the root spans. Probably setting the boolean orphanSpanAsRootSpan as the first parameter resolved this.

I get parentId null for all nodes. And yes I am using startTime as sortCol and desc as sortDir.