genai-telemetry-splunk 2.1.3

GenAI/LLM Observability SDK for Splunk - trace LLM calls, RAG pipelines, and AI agents

GenAI Telemetry for Splunk

GenAI/LLM Observability SDK for Splunk - Trace LLM calls, RAG pipelines, agents, and AI applications with zero-config decorators.

Features

• Zero-config decorators - Add @trace_llm, @trace_chain, etc. to your functions
• Automatic token extraction - Captures input/output tokens from OpenAI, Anthropic, etc.
• Distributed tracing - Links spans across RAG pipelines and agent workflows
• Audit logging - Track user queries, feedback, PII detection, and compliance
• Cost tracking - Monitor token usage and costs per user/model
• Multiple exporters - Splunk HEC, console, file (JSONL)

Installation

pip install genai-telemetry-splunk

With optional dependencies:

# With OpenAI support
pip install genai-telemetry-splunk[openai]

# With Anthropic support
pip install genai-telemetry-splunk[anthropic]

# With all integrations
pip install genai-telemetry-splunk[all]

Quick Start

from genai_telemetry import setup_splunk_telemetry, trace_llm, trace_chain, trace_retrieval
from openai import OpenAI

# 1. Initialize telemetry
setup_splunk_telemetry(
    workflow_name="my-app",
    splunk_hec_url="http://splunk:8088",
    splunk_hec_token="your-token",
    splunk_index="genai_traces"
)

# 2. Create client
client = OpenAI()

# 3. Add decorators
@trace_llm(model_name="gpt-4o-mini", model_provider="openai")
def chat(message: str):
    return client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[{"role": "user", "content": message}]
    )

@trace_retrieval(vector_store="pinecone")
def search(query: str) -> list:
    return [{"text": "doc1"}, {"text": "doc2"}]

@trace_chain(name="rag_pipeline")
def rag(question: str) -> str:
    docs = search(question)
    response = chat(f"Context: {docs}\nQuestion: {question}")
    return response.choices[0].message.content

# 4. Use normally
answer = rag("What is Splunk?")

Decorators

Decorator	Purpose	Key Fields
@trace_llm(model_name, model_provider)	LLM inference calls	input_tokens, output_tokens, duration_ms
@trace_chain(name)	Pipelines (starts new trace)	duration_ms
@trace_retrieval(vector_store)	Vector search	documents_retrieved, duration_ms
@trace_embedding(model)	Embedding generation	duration_ms
@trace_tool(tool_name)	Tool/function calls	duration_ms
@trace_agent(agent_name)	Agent executions (starts new trace)	duration_ms

Configuration Options

setup_splunk_telemetry(
    # Required
    workflow_name="my-app",
    
    # Splunk HEC
    splunk_hec_url="http://splunk:8088",
    splunk_hec_token="your-token",
    splunk_index="genai_traces",
    splunk_sourcetype="genai:trace",
    
    # Optional
    console=False,           # Also print to console
    file_path=None,          # Also write to file (JSONL)
    verify_ssl=False,        # Verify SSL certificates
    batch_size=1,            # Events per batch (1=immediate)
    flush_interval=5.0       # Seconds between flushes
)

OpenAI Example

from genai_telemetry import setup_splunk_telemetry, trace_llm
from openai import OpenAI

setup_splunk_telemetry(
    workflow_name="openai-app",
    splunk_hec_url="http://splunk:8088",
    splunk_hec_token="your-token",
    splunk_index="genai_traces"
)

client = OpenAI()

@trace_llm(model_name="gpt-4o-mini", model_provider="openai")
def chat(message: str):
    return client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[{"role": "user", "content": message}]
    )

response = chat("Hello!")
print(response.choices[0].message.content)

Anthropic Example

from genai_telemetry import setup_splunk_telemetry, trace_llm
from anthropic import Anthropic

setup_splunk_telemetry(
    workflow_name="anthropic-app",
    splunk_hec_url="http://splunk:8088",
    splunk_hec_token="your-token",
    splunk_index="genai_traces"
)

client = Anthropic()

@trace_llm(model_name="claude-3-5-sonnet-20241022", model_provider="anthropic")
def chat(message: str):
    return client.messages.create(
        model="claude-3-5-sonnet-20241022",
        max_tokens=1024,
        messages=[{"role": "user", "content": message}]
    )

response = chat("Hello!")
print(response.content[0].text)

RAG Pipeline Example

from genai_telemetry import (
    setup_splunk_telemetry, 
    trace_llm, 
    trace_chain, 
    trace_retrieval
)
from openai import OpenAI

setup_splunk_telemetry(
    workflow_name="rag-app",
    splunk_hec_url="http://splunk:8088",
    splunk_hec_token="your-token",
    splunk_index="genai_traces"
)

client = OpenAI()

@trace_retrieval(vector_store="pinecone", embedding_model="text-embedding-3-small")
def search(query: str, top_k: int = 3) -> list:
    # Your vector search logic
    return [{"text": "doc1"}, {"text": "doc2"}]

@trace_llm(model_name="gpt-4o-mini", model_provider="openai")
def generate(context: str, question: str):
    return client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[
            {"role": "system", "content": f"Context: {context}"},
            {"role": "user", "content": question}
        ]
    )

@trace_chain(name="rag_pipeline")
def rag(question: str) -> str:
    docs = search(question)
    context = "\n".join([d["text"] for d in docs])
    response = generate(context, question)
    return response.choices[0].message.content

answer = rag("What is Splunk?")

Audit Logging

Track user interactions for compliance:

from genai_telemetry import setup_splunk_telemetry, get_audit_logger

setup_splunk_telemetry(
    workflow_name="my-app",
    splunk_hec_url="http://splunk:8088",
    splunk_hec_token="your-token",
    splunk_index="genai_traces"
)

audit = get_audit_logger()

# Log user query
audit.log_query(user_id="user_123", query="What is AI?")

# Log feedback
audit.log_feedback(user_id="user_123", trace_id="abc", rating=5)

# Log PII detection
audit.log_pii_detection(user_id="user_123", pii_types=["email"], action_taken="masked")

# Log cost
audit.log_cost(user_id="user_123", model_name="gpt-4o-mini", 
               input_tokens=100, output_tokens=50, cost_usd=0.001)

Splunk Queries

# All traces
index=genai_traces

# LLM calls only
index=genai_traces span_type="LLM"

# Errors
index=genai_traces is_error=1

# Token usage by model
index=genai_traces span_type="LLM"
| stats sum(input_tokens) sum(output_tokens) by model_name

# Latency percentiles
index=genai_traces span_type="LLM"
| stats avg(duration_ms) perc95(duration_ms) by model_name

# Audit events
index=genai_audit
| stats count by event_type

Best Practices

Return Full Response for Token Extraction

# Good - returns full response
@trace_llm(model_name="gpt-4o-mini", model_provider="openai")
def chat(message: str):
    return client.chat.completions.create(...)

# Bad - loses token information
@trace_llm(model_name="gpt-4o-mini", model_provider="openai")
def chat(message: str):
    response = client.chat.completions.create(...)
    return response.choices[0].message.content

Use @trace_chain for Entry Points

# Good - chain groups related spans
@trace_chain(name="api_endpoint")
def handle_request(request):
    docs = search(request.query)
    return generate(docs, request.query)

Troubleshooting

No Data in Splunk

1. Check HEC is enabled:

   curl http://splunk:8088/services/collector/health
   

2. Verify token:

   curl -k http://splunk:8088/services/collector/event \
     -H "Authorization: Splunk YOUR_TOKEN" \
     -d '{"event":"test"}'
   

3. Check index exists in Splunk

Zero Token Counts

Return the full response object, not just the content string.

Connection Errors

• Use http:// for non-SSL, https:// for SSL
• Include port: :8088
• Set verify_ssl=False for self-signed certs

License

MIT License - see LICENSE file.

Contributing

Contributions welcome! Please open an issue or PR on GitHub.