|
| 1 | +# Langfuse |
| 2 | + |
| 3 | +[Langfuse](https://langfuse.com) is an open-source LLM engineering platform that provides observability and analytics for AI applications. This integration allows you to **automatically sync LLM metrics from Langfuse into your Mixpanel dashboards**, enabling you to understand how your AI features impact user behavior and business outcomes. |
| 4 | + |
| 5 | +Use this integration to answer questions like: |
| 6 | + |
| 7 | +- _"Are my most active users also the ones who are most engaged with my LLM content?"_ |
| 8 | +- _"Does interacting with the LLM feature relate to higher retention rates?"_ |
| 9 | +- _"How does the LLM feature impact my conversion rates?"_ |
| 10 | +- _"Does the user feedback captured in Langfuse correlate with user behavior in Mixpanel?"_ |
| 11 | + |
| 12 | +## Setup |
| 13 | + |
| 14 | +### Prerequisites |
| 15 | + |
| 16 | +- An active Langfuse account with a project configured |
| 17 | +- Your Mixpanel Project Token (found in Project Settings) |
| 18 | + |
| 19 | +### Configure the integration |
| 20 | + |
| 21 | +1. Log into your Langfuse account and navigate to your project settings |
| 22 | +2. Find the Mixpanel integration section |
| 23 | +3. Select your Mixpanel region: |
| 24 | + - **US**: api.mixpanel.com |
| 25 | + - **EU**: api-eu.mixpanel.com |
| 26 | + - **India**: api-in.mixpanel.com |
| 27 | +4. Enter your Mixpanel Project Token |
| 28 | +5. Enable the integration |
| 29 | + |
| 30 | +### Data synchronization |
| 31 | + |
| 32 | +Once enabled, Langfuse will: |
| 33 | + |
| 34 | +- Perform an initial sync of all historical data from your Langfuse project |
| 35 | +- Automatically sync new data every hour (with a 30-minute delay) |
| 36 | + |
| 37 | +Your Mixpanel dashboards will stay up to date with your latest LLM metrics. |
| 38 | + |
| 39 | +## Data Schema |
| 40 | + |
| 41 | +### User Matching |
| 42 | + |
| 43 | +Langfuse automatically maps user identifiers to ensure seamless data integration: |
| 44 | + |
| 45 | +| Langfuse Field | Mixpanel Field | Description | |
| 46 | +| ------------------------------------- | -------------- | ------------------------------------------------------------------------------------------- | |
| 47 | +| `user_id` | `distinct_id` | Primary user identifier | |
| 48 | +| Trace/generation/score timestamp | `time` | Event timestamp (milliseconds since epoch) | |
| 49 | +| `trace.metadata.$mixpanel_session_id` | `session_id` | Optional session identifier (add this to your Langfuse trace metadata for session tracking) | |
| 50 | + |
| 51 | +### Events |
| 52 | + |
| 53 | +The integration sends three event types to Mixpanel: |
| 54 | + |
| 55 | +#### `[Langfuse] Trace` |
| 56 | + |
| 57 | +Represents a complete LLM interaction (e.g., a user conversation or workflow). |
| 58 | + |
| 59 | +**Properties:** |
| 60 | + |
| 61 | +- `time`: Milliseconds since epoch when the event occurred |
| 62 | +- `distinct_id`: User ID or anonymous identifier |
| 63 | +- `$user_id`: User ID sent to Mixpanel's native user ID field |
| 64 | +- `$insert_id`: Unique identifier for deduplication |
| 65 | +- `session_id`: Optional session identifier (from `$mixpanel_session_id` in metadata, or falls back to Langfuse session_id) |
| 66 | +- `langfuse_trace_name`: The name of the trace |
| 67 | +- `langfuse_url`: The URL of the trace in Langfuse |
| 68 | +- `langfuse_user_url`: Deep link to the user profile in Langfuse |
| 69 | +- `langfuse_id`: The unique identifier of the trace |
| 70 | +- `langfuse_cost_usd`: The total cost associated with the trace |
| 71 | +- `langfuse_count_observations`: The number of observations (LLM calls) in the trace |
| 72 | +- `langfuse_session_id`: The session ID related to the event |
| 73 | +- `langfuse_project_id`: The project ID associated with the event |
| 74 | +- `langfuse_user_id`: User ID related to the event (defaults to `langfuse_unknown_user` if null) |
| 75 | +- `langfuse_latency`: The latency of the trace in milliseconds |
| 76 | +- `langfuse_release`: Release information associated with the trace |
| 77 | +- `langfuse_version`: The version of the trace |
| 78 | +- `langfuse_tags`: Tags associated with the trace |
| 79 | +- `langfuse_environment`: The environment associated with the trace (e.g., production, staging) |
| 80 | +- `langfuse_event_version`: The integration version of Langfuse |
| 81 | + |
| 82 | +#### `[Langfuse] Generation` |
| 83 | + |
| 84 | +Represents an individual LLM generation (e.g., a single API call to OpenAI, Anthropic, etc.). |
| 85 | + |
| 86 | +**Properties:** |
| 87 | + |
| 88 | +- `time`: Milliseconds since epoch when the generation started |
| 89 | +- `distinct_id`: User ID or anonymous identifier |
| 90 | +- `$user_id`: User ID sent to Mixpanel's native user ID field |
| 91 | +- `$insert_id`: Unique identifier for deduplication |
| 92 | +- `session_id`: Optional session identifier (from `$mixpanel_session_id` in metadata, or falls back to Langfuse session_id) |
| 93 | +- `langfuse_generation_name`: The name of the generation |
| 94 | +- `langfuse_trace_name`: Name of the trace related to the generation |
| 95 | +- `langfuse_trace_id`: The unique identifier of the trace related to the generation |
| 96 | +- `langfuse_url`: The URL of the generation in Langfuse |
| 97 | +- `langfuse_user_url`: Deep link to the user profile in Langfuse |
| 98 | +- `langfuse_id`: Unique identifier of the generation |
| 99 | +- `langfuse_cost_usd`: Computed total cost of the generation |
| 100 | +- `langfuse_input_units`: Number of tokens used in the input/prompt |
| 101 | +- `langfuse_output_units`: Number of tokens produced by the generation |
| 102 | +- `langfuse_total_units`: Total number of tokens consumed in the generation process |
| 103 | +- `langfuse_session_id`: The session ID associated with the trace of the generation |
| 104 | +- `langfuse_project_id`: The project ID where the generation occurred |
| 105 | +- `langfuse_user_id`: The user ID that started the trace linked to the generation (defaults to `langfuse_unknown_user` if unavailable) |
| 106 | +- `langfuse_latency`: The observed latency of the generation in milliseconds |
| 107 | +- `langfuse_time_to_first_token`: The time taken to generate the first token when streaming (milliseconds) |
| 108 | +- `langfuse_release`: Release information of the trace attached to the generation |
| 109 | +- `langfuse_version`: The version information about the generation |
| 110 | +- `langfuse_model`: The model used during this generation (e.g., gpt-4, claude-3-sonnet) |
| 111 | +- `langfuse_level`: The level associated with the generation |
| 112 | +- `langfuse_tags`: Tags attached to the trace of the generation |
| 113 | +- `langfuse_environment`: The environment associated with the generation |
| 114 | +- `langfuse_event_version`: The integration version with Langfuse |
| 115 | + |
| 116 | +#### `[Langfuse] Score` |
| 117 | + |
| 118 | +Represents user feedback, evaluations, or quality metrics. |
| 119 | + |
| 120 | +**Properties:** |
| 121 | + |
| 122 | +- `time`: Milliseconds since epoch when the score event occurred |
| 123 | +- `distinct_id`: User ID or anonymous identifier |
| 124 | +- `$user_id`: User ID sent to Mixpanel's native user ID field |
| 125 | +- `$insert_id`: Unique identifier for deduplication |
| 126 | +- `session_id`: Optional session identifier (from `$mixpanel_session_id` in metadata, or falls back to Langfuse session_id) |
| 127 | +- `langfuse_score_name`: The name associated with the score (e.g., "user_feedback", "accuracy") |
| 128 | +- `langfuse_score_value`: The numeric value of the score |
| 129 | +- `langfuse_score_string_value`: The string value of the score (for BOOLEAN and CATEGORICAL scores) |
| 130 | +- `langfuse_score_data_type`: The data type of the score (NUMERIC, BOOLEAN, or CATEGORICAL) |
| 131 | +- `langfuse_score_comment`: Comments attached to the score |
| 132 | +- `langfuse_score_metadata`: Additional metadata attached to the score |
| 133 | +- `langfuse_trace_name`: The name of the trace associated with the score |
| 134 | +- `langfuse_trace_id`: The unique identifier of the trace associated with the score |
| 135 | +- `langfuse_user_url`: Deep link to the user profile in Langfuse |
| 136 | +- `langfuse_id`: The unique identifier of the score |
| 137 | +- `langfuse_session_id`: The session ID related to the score's trace |
| 138 | +- `langfuse_project_id`: The project ID linked with the score's trace |
| 139 | +- `langfuse_user_id`: The user ID that triggered the trace tied to the score (defaults to `langfuse_unknown_user` if not available) |
| 140 | +- `langfuse_release`: The release information of the trace associated with the score |
| 141 | +- `langfuse_tags`: Tags related to the trace of the score |
| 142 | +- `langfuse_environment`: The environment associated with the score |
| 143 | +- `langfuse_event_version`: The integration version with Langfuse |
| 144 | + |
| 145 | +## Use Cases |
| 146 | + |
| 147 | +### Get Started with the Analytics for AI Dashboard Template |
| 148 | + |
| 149 | +The fastest way to see value from this integration is to use Mixpanel's **Analytics for AI dashboard template**. This pre-built dashboard provides instant insights into how your LLM features are performing and how they impact user behavior. |
| 150 | + |
| 151 | +[**View the Analytics for AI Dashboard Template →**](https://mixpanel.com/p/NaKPyubj6EuA4oV75taqrq) |
| 152 | + |
| 153 | +The template includes ready-to-use reports for: |
| 154 | + |
| 155 | +- **LLM Feature Adoption**: Track how many users are engaging with your AI features |
| 156 | +- **Cost Analysis**: Monitor your LLM spending by user and feature |
| 157 | +- **Performance Metrics**: Visualize latency, token usage, and generation times |
| 158 | +- **User Feedback**: Analyze scores and ratings from Langfuse |
| 159 | +- **Retention Impact**: Understand retention rates of AI feature users |
| 160 | + |
| 161 | +### Analyze LLM Feature Adoption |
| 162 | + |
| 163 | +Create funnels to track: |
| 164 | + |
| 165 | +- Users who trigger `[Langfuse] Trace` events |
| 166 | +- Conversion to key actions in your product |
| 167 | +- Retention rates for AI feature users vs. non-users |
| 168 | + |
| 169 | +### Monitor LLM Costs by User Segment |
| 170 | + |
| 171 | +Build insights to: |
| 172 | + |
| 173 | +- Group users by `langfuse_cost_usd` total spend |
| 174 | +- Segment by user properties (plan type, company size, etc.) |
| 175 | +- Identify high-cost users or sessions |
| 176 | + |
| 177 | +### Correlate User Feedback with Behavior |
| 178 | + |
| 179 | +Analyze how `[Langfuse] Score` events relate to: |
| 180 | + |
| 181 | +- Session length and engagement |
| 182 | +- Feature usage patterns |
| 183 | +- Churn or upgrade likelihood |
| 184 | + |
| 185 | +### Track Model Performance Impact |
| 186 | + |
| 187 | +Compare: |
| 188 | + |
| 189 | +- `langfuse_latency` across different `langfuse_model` values |
| 190 | +- Token usage efficiency (`langfuse_total_units`) |
| 191 | +- Cost differences between model versions |
| 192 | + |
| 193 | +## Troubleshooting |
| 194 | + |
| 195 | +**Events not appearing in Mixpanel?** |
| 196 | + |
| 197 | +- Verify you selected the correct Mixpanel region in Langfuse |
| 198 | +- Confirm your Project Token is correct |
| 199 | +- Allow up to 90 minutes for the first sync to complete |
| 200 | +- Check that your Langfuse project has trace data |
| 201 | + |
| 202 | +**User matching issues?** |
| 203 | + |
| 204 | +- Ensure the `user_id` in Langfuse matches the `distinct_id` in Mixpanel |
| 205 | +- For session tracking, add `$mixpanel_session_id` to your Langfuse trace metadata |
| 206 | + |
| 207 | +**Need additional help?** |
| 208 | +Contact Langfuse support or submit a feature request on their [ideas board](https://langfuse.com/ideas). |
| 209 | + |
| 210 | +## Learn More |
| 211 | + |
| 212 | +- [Langfuse Mixpanel Integration Documentation](https://langfuse.com/integrations/analytics/mixpanel) |
| 213 | +- [Langfuse Documentation](https://langfuse.com/docs) |
| 214 | +- [Langfuse GitHub](https://github.com/langfuse/langfuse) |
| 215 | +- [LLM Observability Best Practices](https://langfuse.com/docs/observability) |
0 commit comments