Agentic Workflow Service

Overview

The Agentic Workflow Service is a production-deployed, Go-based backend that demonstrates how to build an agentic AI system capable of:

• Understanding user intent
• Orchestrating multiple AI tools
• Coordinating retrieval and summarization microservices
• Evaluating results for grounding
• Producing structured, traceable outputs

This service acts as an agent orchestration layer, coordinating multiple internal AI microservices rather than embedding all AI logic in a single endpoint.

It is designed to reflect how real production AI backends are built: modular, observable, explainable, and incrementally extensible.

---

Why This Service Exists

Many AI backends either:

• Call an LLM directly with no structure, or
• Hard-code a single RAG flow without reasoning, validation, or orchestration

This service demonstrates a production-minded agentic approach that is:

• Simple enough to understand end-to-end
• Structured enough to support real-world AI workflows

Specifically, it showcases:

• Explicit agent workflows
• Tool orchestration across services
• Grounding and evaluation checks
• Request-level tracing
• Service-to-service AI architecture

---

High-Level Architecture

Client
  |
  | POST /run
  v
Agentic Workflow Service
  |
  |-- LLM (intent detection)
  |-- Tool: Semantic Search (RAG Notes API)
  |-- Tool: Summarization (AI Summary Service)
  |-- LLM (final answer synthesis)
  |-- Evaluation (grounding check)

Each request is treated as an agent run, with its own execution context and trace ID.

---

Agent Execution Flow

Each /run request follows a deterministic, inspectable workflow:

1. Intent Detection (LLM) Determines what the user is trying to do and extracts entities (e.g. topic).

2. Retrieval (Semantic Search Tool) Executes semantic search against an external RAG Notes API to retrieve relevant information.

3. Summarization (Summary Tool) Passes retrieved context to a dedicated AI Summary microservice to condense and normalize information before reasoning.

4. Final Answer Synthesis (LLM) Generates a grounded, structured response using summarized and retrieved context.

5. Evaluation Scores the output for grounding and observability.

The workflow is explicit and readable to keep reasoning, debugging, and extension straightforward.

---

Workflow Definition

The agent workflow is defined as an ordered sequence of steps:

var Workflow = []Step{
    {Type: StepLLM, Prompt: "intent_decider"},
    {Type: StepTool, Tool: "search"},
    {Type: StepTool, Tool: "summarize"},
    {Type: StepLLM, Prompt: "final_answer"},
    {Type: StepEvaluate},
}

This design makes the agent:

• Deterministic
• Inspectable
• Easy to extend (additional tools, retries, or conditional logic)

---

Tools

Semantic Search Tool (RAG Notes API)

The agent integrates with an external RAG Notes API that provides:

• Embedding-based semantic search
• Top-K retrieval of relevant notes

The agent does not rely on keyword search. Instead, it retrieves candidate information semantically and reasons over the results afterward.

---

Summarization Tool (AI Summary Service)

Retrieved context is passed to a dedicated AI Summary microservice, which:

• Condenses multiple retrieved documents
• Normalizes noisy or overlapping information
• Produces a focused summary for downstream reasoning

This keeps the agent lightweight and avoids coupling summarization logic directly into the orchestration layer.

---

Evaluation & Grounding

Before producing a final response, the agent runs an evaluation step that:

• Inspects retrieved and summarized context
• Scores how well the final answer is grounded in evidence
• Emits an evaluation score for observability

This reflects a production mindset where LLM outputs are treated as unreliable by default and must be validated.

---

Observability & Tracing

Each agent run is assigned a unique run_id.

The service logs:

• Step-level execution
• Tool invocations
• Evaluation outcomes

Example log output:

[agent] run=8e29f3fb... step_type=llm step=intent_decider
[agent] run=8e29f3fb... step_type=tool step=search
[agent] run=8e29f3fb... step_type=tool step=summarize
[agent] run=8e29f3fb... step_type=llm step=final_answer
[agent] run=8e29f3fb... evaluation_score=0.60 pass=true

This allows each request to be traced and debugged deterministically.

---

API

POST /run

Runs the agent workflow for a single input.

Request

{
  "input": "What notes do I have about Docker?"
}

Response

{
  "status": "ok",
  "run_id": "8e29f3fb-c38e-41dd-a560-9a9f844cb9aa",
  "result": {
    "intent_decider": { ... },
    "search": { ... },
    "summarize": { ... },
    "evaluation_score": 0.6,
    "final_answer": { ... }
  }
}

---

What This Project Demonstrates

• Agentic AI workflows
• Tool orchestration across microservices
• RAG-based retrieval
• Summarization as a first-class tool
• Grounding and evaluation
• Lightweight observability
• Clean, production-oriented Go backend design

This service is intentionally focused and composable, designed to orchestrate AI capabilities, not replace them.

---

Tech Stack

• Go
• Fiber
• Docker
• Fly.io
• OpenAI API
• External RAG & Summarization services
• JSON-based agent state

---

Final Note

This project prioritizes clarity, correctness, and explainability over complexity. It reflects how real AI backend systems evol