Instrument LLM Analytics

AI agents: this is one page from PostHog's docs. Full index of Markdown docs for LLMs: https://posthog.com/llms.txt

Manual capture AI Observability installation

Canonical URL: https://posthog.com/docs/ai-observability/installation/manual-capture

Cloud region: This page uses US Cloud hosts. For an EU Cloud project, use https://eu.i.posthog.com instead of https://us.i.posthog.com (ingestion host), and use https://eu.posthog.com instead of https://app.posthog.com (app host for links).

On this page

• Capture LLM events manually

• Event properties

• Verify traces and generations

• Next steps

• Large events

Image: untitled

Image: untitled

Let AI instrument your LLM calls for you

Skip the manual setup — run this in your project and the wizard installs the SDK and wires up AI Observability for you.

npx @posthog/wizard ai-observability

Image: PostHog Wizard hedgehog

Image: untitledLet AI instrument your LLM calls for you

1. 1

   Capture LLM events manually

   If you use a different server-side SDK, or prefer to use the API, capture the data manually. Call the capture method, or use the capture API.

   $ai_trace_id groups the events of one LLM interaction into a trace. $ai_session_id is optional and groups related traces into a session, which is what the Sessions tab reads. Set it if your product has multi-turn conversations. Workloads that finish in a single trace, like batch jobs or one-shot generation, do not need it. Send it as null on those to say so explicitly, which tells the instrumentation checklist the workload is complete rather than missing a session id.

   API

   Capture via API

   Copy

   curl -X POST "https://us.i.posthog.com/i/v0/e/" \
           -H "Content-Type: application/json" \
           -d '{
               "api_key": "<ph_project_token>",
               "event": "$ai_generation",
               "properties": {
                   "distinct_id": "user_123",
                   "$ai_trace_id": "trace_id_here",
                   "$ai_session_id": "conversation-abc",
                   "$ai_model": "gpt-5-mini",
                   "$ai_provider": "openai",
                   "$ai_input": [{"role": "user", "content": "Tell me a fun fact about hedgehogs"}],
                   "$ai_input_tokens": 10,
                   "$ai_output_choices": [{"role": "assistant", "content": "Hedgehogs have around 5,000 to 7,000 spines on their backs!"}],
                   "$ai_output_tokens": 20,
                   "$ai_latency": 1.5,
                   "$ai_stream": true,
                   "$ai_time_to_first_token": 0.25
               }
           }'

   Node.js

   1. Install

   Copy

   npm install posthog-node

   2. Initialize PostHog

   Copy

   import { PostHog } from 'posthog-node'
   
   const client = new PostHog('<ph_project_token>', {
       host: 'https://us.i.posthog.com'
   })

   3. Capture Event

   Copy

   // After your LLM call
   client.capture({
       distinctId: 'user_123',
       event: '$ai_generation',
       properties: {
           $ai_trace_id: 'trace_id_here',
           $ai_session_id: 'conversation-abc', // optional: groups traces into one session
           $ai_model: 'gpt-5-mini',
           $ai_provider: 'openai',
           $ai_input: [{ role: 'user', content: 'Tell me a fun fact about hedgehogs' }],
           $ai_input_tokens: 10,
           $ai_output_choices: [{ role: 'assistant', content: 'Hedgehogs have around 5,000 to 7,000 spines on their backs!' }],
           $ai_output_tokens: 20,
           $ai_latency: 1.5,
           // For streaming responses, also include:
           // $ai_stream: true,
           // $ai_time_to_first_token: 0.25
       }
   })
   
   client.shutdown()

   Python

   1. Install

   Copy

   pip install posthog

   2. Initialize PostHog

   Copy

   from posthog import Posthog
   
   posthog = Posthog("<ph_project_token>", host="https://us.i.posthog.com")

   3. Capture Event

   Copy

   # After your LLM call
   posthog.capture(
       distinct_id='user_123',
       event='$ai_generation',
       properties={
           '$ai_trace_id': 'trace_id_here',
           '$ai_session_id': 'conversation-abc',  # optional: groups traces into one session
           '$ai_model': 'gpt-5-mini',
           '$ai_provider': 'openai',
           '$ai_input': [{'role': 'user', 'content': 'Tell me a fun fact about hedgehogs'}],
           '$ai_input_tokens': 10,
           '$ai_output_choices': [{'role': 'assistant', 'content': 'Hedgehogs have around 5,000 to 7,000 spines on their backs!'}],
           '$ai_output_tokens': 20,
           '$ai_latency': 1.5,
           # For streaming responses, also include:
           # '$ai_stream': True,
           # '$ai_time_to_first_token': 0.25
       }
   )

   Go

   1. Install

   Copy

   go get github.com/posthog/posthog-go

   2. Initialize PostHog

   Copy

   import "github.com/posthog/posthog-go"
   
   client, _ := posthog.NewWithConfig("<ph_project_token>", posthog.Config{
       Endpoint: "https://us.i.posthog.com",
   })
   defer client.Close()

   3. Capture Event

   Copy

   // After your LLM call
   client.Enqueue(posthog.Capture{
       DistinctId: "user_123",
       Event:      "$ai_generation",
       Properties: map[string]interface{}{
           "$ai_trace_id":        "trace_id_here",
           // $ai_session_id is optional: it groups traces into one session
           "$ai_session_id":      "conversation-abc",
           "$ai_model":           "gpt-5-mini",
           "$ai_provider":        "openai",
           "$ai_input_tokens":    10,
           "$ai_output_tokens":   20,
           "$ai_latency":         1.5,
           // For streaming responses, also include:
           // "$ai_stream":              true,
           // "$ai_time_to_first_token": 0.25,
       },
   })

   Ruby

   1. Install

   Copy

   gem install posthog-ruby

   2. Initialize PostHog

   Copy

   require 'posthog'
   
   posthog = PostHog::Client.new({
       api_key: '<ph_project_token>',
       host: 'https://us.i.posthog.com'
   })

   3. Capture Event

   Copy

   # After your LLM call
   posthog.capture({
       distinct_id: 'user_123',
       event: '$ai_generation',
       properties: {
       '$ai_trace_id' => 'trace_id_here',
       '$ai_session_id' => 'conversation-abc', # optional: groups traces into one session
       '$ai_model' => 'gpt-5-mini',
       '$ai_provider' => 'openai',
       '$ai_input_tokens' => 10,
       '$ai_output_tokens' => 20,
       '$ai_latency' => 1.5
       # For streaming responses, also include:
       # '$ai_stream' => true,
       # '$ai_time_to_first_token' => 0.25
       }
   })

   PHP

   1. Install

   Copy

   composer require posthog/posthog-php

   2. Initialize PostHog

   Copy

   <?php
   require_once __DIR__ . '/vendor/autoload.php';
   use PostHog\PostHog;
   
   PostHog::init('<ph_project_token>', [
       'host' => 'https://us.i.posthog.com'
   ]);

   3. Capture Event

   Copy

   // After your LLM call
   PostHog::capture([
       'distinctId' => 'user_123',
       'event' => '$ai_generation',
       'properties' => [
           '$ai_trace_id' => 'trace_id_here',
           '$ai_session_id' => 'conversation-abc', // optional: groups traces into one session
           '$ai_model' => 'gpt-5-mini',
           '$ai_provider' => 'openai',
           '$ai_input_tokens' => 10,
           '$ai_output_tokens' => 20,
           '$ai_latency' => 1.5
           // For streaming responses, also include:
           // '$ai_stream' => true,
           // '$ai_time_to_first_token' => 0.25
       ]
   ]);

2. 2

   Event properties

   Each event type has specific properties. See the tabs below for detailed property documentation for each event type.

   Generation

   A generation is a single call to an LLM.

   Event name: $ai_generation

   Core properties

   Property	Description
   $ai_trace_id	The trace ID (a UUID to group AI events) like conversation_id Must contain only letters, numbers, and special characters: -, _, ~, ., @, (, ), !, ', :, | Example: d9222e05-8708-41b8-98ea-d4a21849e761
   $ai_session_id	(Optional) Groups related traces into a session, which is what the Sessions tab reads. Set it if your product has multi-turn conversations. A workload that finishes in a single trace does not need it. Send it as null to say so explicitly, which tells the instrumentation checklist the workload is complete rather than missing a session id. Example: session-abc-123, conv-user-456
   $ai_span_id	(Optional) Unique identifier for this generation
   $ai_span_name	(Optional) Name given to this generation Example: summarize_text
   $ai_parent_id	(Optional) Parent span ID for tree view grouping
   $ai_model	The model used Example: gpt-5-mini
   $ai_provider	The LLM provider Example: openai, anthropic, gemini
   $ai_input	List of messages sent to the LLM. Each message should have a role property with one of: "user", "system", or "assistant" Example: json [ { "role": "user", "content": [ { "type": "text", "text": "What's in this image?" }, { "type": "image", "image": "https://example.com/image.jpg" }, { "type": "function", "function": { "name": "get_weather", "arguments": { "location": "San Francisco" } } } ] } ]
   $ai_input_tokens	The number of tokens in the input (often found in response.usage)
   $ai_output_choices	List of response choices from the LLM. Each choice should have a role property with one of: "user", "system", or "assistant" Example: json [ { "role": "assistant", "content": [ { "type": "text", "text": "I can see a hedgehog in the image." }, { "type": "function", "function": { "name": "get_weather", "arguments": { "location": "San Francisco" } } } ] } ]
   $ai_output_tokens	The number of tokens in the output (often found in response.usage)
   $ai_latency	(Optional) The latency of the LLM call in seconds
   $ai_time_to_first_token	(Optional) Time to first token in seconds. Only applicable for streaming responses.
   $ai_http_status	(Optional) The HTTP status code of the response
   $ai_base_url	(Optional) The base URL of the LLM provider Example: https://api.openai.com/v1
   $ai_request_url	(Optional) The full URL of the request made to the LLM API Example: https://api.openai.com/v1/chat/completions
   $ai_is_error	(Optional) Boolean to indicate if the request was an error
   $ai_error	(Optional) The error message or object
   $ai_stop_reason	(Optional) The reason the model stopped generating tokens Example: end_turn, stop, max_tokens, tool_use

   Cost properties

   Cost properties are optional as we can automatically calculate them from model and token counts. If you want, you can provide your own cost properties or custom pricing instead.

   Pre-calculated costs

   Property	Description
   $ai_input_cost_usd	(Optional) The cost in USD of the input tokens
   $ai_output_cost_usd	(Optional) The cost in USD of the output tokens
   $ai_request_cost_usd	(Optional) The cost in USD for the requests
   $ai_web_search_cost_usd	(Optional) The cost in USD for the web searches
   $ai_total_cost_usd	(Optional) The total cost in USD (sum of all cost components)
   $ai_cost_passthrough	(Optional) Set this when your provider reports the real cost, such as an LLM gateway. We keep your $ai_total_cost_usd and leave the input and output costs unset.

   Custom pricing

   Property	Description
   $ai_input_token_price	(Optional) Price per input token (used to calculate $ai_input_cost_usd)
   $ai_output_token_price	(Optional) Price per output token (used to calculate $ai_output_cost_usd)
   $ai_cache_read_token_price	(Optional) Price per cached token read
   $ai_cache_write_token_price	(Optional) Price per cached token write. For custom Anthropic pricing, this applies to both cache TTLs unless $ai_cache_write_1h_token_price is set.
   $ai_cache_write_1h_token_price	(Optional) Price per token written to Anthropic's 1-hour cache. Takes precedence over $ai_cache_write_token_price for 1-hour writes.
   $ai_request_price	(Optional) Price per request
   $ai_request_count	(Optional) Number of requests (defaults to 1 if $ai_request_price is set)
   $ai_web_search_price	(Optional) Price per web search
   $ai_web_search_count	(Optional) Number of web searches performed

   Cache properties

   Property	Description
   $ai_cache_read_input_tokens	(Optional) Number of tokens read from cache
   $ai_cache_creation_input_tokens	(Optional) Number of tokens written to cache For Anthropic events, PostHog uses the TTL-specific counts instead of this aggregate when both are present. The aggregate should equal their sum. If either count is missing, PostHog uses the aggregate. Built-in Gemini pricing uses normal input rates for cache writes and does not estimate cache storage fees. Custom pricing can override these rates.
   $ai_cache_creation_5m_input_tokens	(Optional) Number of tokens written to Anthropic's 5-minute cache
   $ai_cache_creation_1h_input_tokens	(Optional) Number of tokens written to Anthropic's 1-hour cache
   $ai_cache_reporting_exclusive	(Optional) Whether cache tokens are excluded from $ai_input_tokens. When true, cache tokens are separate from input tokens. When false, input tokens already include cache tokens. Defaults to true for Anthropic provider or Claude models, false otherwise.

   Model parameters

   Property	Description
   $ai_temperature	(Optional) Temperature parameter used in the LLM request
   $ai_stream	(Optional) Whether the response was streamed
   $ai_max_tokens	(Optional) Maximum tokens setting for the LLM response
   $ai_tools	(Optional) Tools/functions available to the LLM Example: json [ { "type": "function", "function": { "name": "get_weather", "parameters": {} } } ]

   Trace

   A trace is a group that contains multiple spans, generations, and embeddings. Traces can be manually sent as events or appear as pseudo-events automatically created from child events.

   Event name: $ai_trace

   Core properties

   Property	Description
   $ai_trace_id	The trace ID (a UUID to group related AI events together) Must contain only letters, numbers, and special characters: -, _, ~, ., @, (, ), !, ', :, | Example: d9222e05-8708-41b8-98ea-d4a21849e761
   $ai_session_id	(Optional) Groups related traces into a session, which is what the Sessions tab reads. Set it if your product has multi-turn conversations. A workload that finishes in a single trace does not need it. Send it as null to say so explicitly, which tells the instrumentation checklist the workload is complete rather than missing a session id. Must contain only letters, numbers, and special characters: -, _, ~, ., @, (, ), !, ', :, | Example: session-abc-123, conv-user-456
   $ai_input_state	The input of the whole trace Example: json [ { "role": "user", "content": "What's the weather in SF?" } ] or any JSON-serializable state
   $ai_output_state	The output of the whole trace Example: json [ { "role": "assistant", "content": "The weather in San Francisco is..." } ] or any JSON-serializable state
   $ai_latency	(Optional) The latency of the trace in seconds
   $ai_span_name	(Optional) The name of the trace Example: chat_completion, rag_pipeline
   $ai_is_error	(Optional) Boolean to indicate if the trace encountered an error
   $ai_error	(Optional) The error message or object if the trace failed

   Pseudo-trace Events

   When you send generation ($ai_generation), span ($ai_span), or embedding ($ai_embedding) events with a $ai_trace_id, PostHog automatically creates a pseudo-trace event that appears in the dashboard as a parent grouping. These pseudo-traces:

   • Are not actual events in your data

   • Automatically aggregate metrics from child events (latency, tokens, costs)

   • Provide a hierarchical view of your AI operations

   • Do not require sending an explicit $ai_trace event

   This means you can either:

   1. Send explicit $ai_trace events to control the trace metadata

   2. Let PostHog automatically create pseudo-traces from your generation/span events

   Span

   A span is a single action within your application, such as a function call or vector database search.

   Event name: $ai_span

   Core properties

   Property	Description
   $ai_trace_id	The trace ID (a UUID to group related AI events together) Must contain only letters, numbers, and the following characters: -, _, ~, ., @, (, ), !, ', :, | Example: d9222e05-8708-41b8-98ea-d4a21849e761
   $ai_session_id	(Optional) Groups related traces into a session, which is what the Sessions tab reads. Set it if your product has multi-turn conversations. A workload that finishes in a single trace does not need it. Send it as null to say so explicitly, which tells the instrumentation checklist the workload is complete rather than missing a session id. Example: session-abc-123, conv-user-456
   $ai_span_id	(Optional) Unique identifier for this span Example: bdf42359-9364-4db7-8958-c001f28c9255
   $ai_span_name	(Optional) The name of the span Example: vector_search, data_retrieval, tool_call
   $ai_parent_id	(Optional) Parent ID for tree view grouping (trace_id or another span_id) Example: 537b7988-0186-494f-a313-77a5a8f7db26
   $ai_input_state	The input state of the span Example: json { "query": "search for documents about hedgehogs" } or any JSON-serializable state
   $ai_output_state	The output state of the span Example: json { "results": ["doc1", "doc2"], "count": 2 } or any JSON-serializable state
   $ai_latency	(Optional) The latency of the span in seconds Example: 0.361
   $ai_is_error	(Optional) Boolean to indicate if the span encountered an error
   $ai_error	(Optional) The error message or object if the span failed Example: json { "message": "Connection timeout", "code": "TIMEOUT" }

   Embedding

   An embedding is a single call to an embedding model to convert text into a vector representation.

   Event name: $ai_embedding

   Core properties

   Property	Description
   $ai_trace_id	The trace ID (a UUID to group related AI events together). Must contain only letters, numbers, and special characters: -, _, ~, ., @, (, ), !, ', :, | Example: d9222e05-8708-41b8-98ea-d4a21849e761
   $ai_session_id	(Optional) Groups related traces into a session, which is what the Sessions tab reads. Set it if your product has multi-turn conversations. A workload that finishes in a single trace does not need it. Send it as null to say so explicitly, which tells the instrumentation checklist the workload is complete rather than missing a session id. Example: session-abc-123, conv-user-456
   $ai_span_id	(Optional) Unique identifier for this embedding operation
   $ai_span_name	(Optional) Name given to this embedding operation Example: embed_user_query, index_document
   $ai_parent_id	(Optional) Parent span ID for tree-view grouping
   $ai_model	The embedding model used Example: text-embedding-3-small, text-embedding-ada-002
   $ai_provider	The LLM provider Example: openai, cohere, voyage
   $ai_input	The text to embed Example: "Tell me a fun fact about hedgehogs" or array of strings for batch embeddings
   $ai_input_tokens	The number of tokens in the input
   $ai_latency	(Optional) The latency of the embedding call in seconds
   $ai_http_status	(Optional) The HTTP status code of the response
   $ai_base_url	(Optional) The base URL of the LLM provider Example: https://api.openai.com/v1
   $ai_request_url	(Optional) The full URL of the request made to the embedding API Example: https://api.openai.com/v1/embeddings
   $ai_is_error	(Optional) Boolean to indicate if the request was an error
   $ai_error	(Optional) The error message or object if the embedding failed

   Cost properties

   Cost properties are optional as we can automatically calculate them from model and token counts. If you want, you can provide your own cost property instead.

   Property	Description
   $ai_input_cost_usd	(Optional) Cost in USD for input tokens
   $ai_output_cost_usd	(Optional) Cost in USD for output tokens (usually 0 for embeddings)
   $ai_total_cost_usd	(Optional) Total cost in USD

3. Verify traces and generations

   Recommended

   Confirm LLM events are being sent to PostHog

   Let's make sure LLM events are being captured and sent to PostHog. Under AI Observability, you should see rows of data appear in the Traces and Generations tabs.

   Image: LLM generations in PostHog

   Check for LLM events in PostHog

4. 3

   Next steps

   Recommended

   Now that you're capturing AI conversations, continue with the resources below to learn what else AI Observability enables within the PostHog platform.

   Resource	Description
   Basics	Learn the basics of how LLM calls become events in PostHog.
   Generations	Read about the $ai_generation event and its properties.
   Traces	Explore the trace hierarchy and how to use it to debug LLM calls.
   Spans	Review spans and their role in representing individual operations.
   Anaylze LLM performance	Learn how to create dashboards to analyze LLM performance.

Large events

For large events, use the dedicated AI ingestion path – see capturing large AI events.

In Python, capture() drops events over 900 KiB. Call capture_ai() instead, with the same arguments, to send events up to 8MB:

posthog.capture_ai("$ai_generation", distinct_id="user_123", properties={...})