Sentinel Research Agent (SRA)

Why This Exists

This project explores how AI-assisted workflows and orchestration systems can automate repetitive research and operational tasks using structured multi-step processes.

SRA takes a natural-language question, runs an iterative research workflow, and returns a validated JSON report with sources. The focus is workflow design, orchestration reliability, and structured outputs that can plug into downstream operational tooling.

Workflow At a Glance

flowchart TD
    A[User Query] --> B[Planner]
    B --> C[Search Tool\nGoogle Custom Search]
    C --> D[Analyzer]
    D -->|CONTINUE| C
    D -->|FINISH| E[Reporter]
    D -->|max_iters reached| E
    E --> F[FinalReport JSON]

Loading

flowchart LR
    U[CLI / UI Input] --> G[LangGraph Orchestrator]
    G --> P[Planner Model]
    G --> S[Search Node]
    S --> GS[Google Search API]
    G --> A[Analyzer Model]
    G --> R[Reporter Model]
    R --> V[Pydantic Validation\nFinalReport]
    V --> O[JSON Output]

    C[(Config / Env)] --> G
    C --> P
    C --> A
    C --> R

Loading

UI Preview

Research In Progress

Completed Report

What It Does

• Runs a planner/search/analyzer loop to gather evidence.
• Uses Google Custom Search for external retrieval.
• Produces a validated FinalReport object (topic, executive_summary, sections, sources).
• Supports CLI and web UI execution paths.

Practical Use Cases

• Lead qualification research before sales outreach.
• Research automation for market, policy, and competitor scans.
• Internal reporting briefs with source-linked summaries.
• Workflow routing inputs for downstream automation.
• CRM enrichment support using externally sourced context.
• Operational dashboards fed by structured JSON outputs.

Current Limitations

This project is experimental and currently focused on orchestration logic and workflow design rather than production deployment.

• Depends on third-party model and search APIs, so latency and availability vary.
• Retrieval quality depends on search index coverage and ranking.
• Structured output fallbacks are robust but still model-behavior dependent.
• No built-in persistence, job queueing, or multi-user auth layer.
• Not tuned yet for high-throughput production workloads.

Quick Start

1. Install

python -m venv .venv
source .venv/bin/activate
pip install -e .

2. Configure Environment

OpenRouter is the default provider.

export LLM_API_KEY=...
export LLM_MODEL=google/gemini-2.5-flash
export LLM_BASE_URL=https://openrouter.ai/api/v1

export GOOGLE_SEARCH_API_KEY=...
export GOOGLE_SEARCH_ENGINE_ID=...

You can also place these values in .env.

3. Run

Both forms are supported:

sra "How is the EU regulating frontier AI safety tests?"

sra run "How is the EU regulating frontier AI safety tests?"

Control loop depth:

sra run "query" --max-iters 4

Run the web UI:

sra ui --host 127.0.0.1 --port 8000

Terminal Example

sra run "What changed in frontier AI safety requirements across EU, UK, and US in the last 12 months?" --max-iters 2

Expected output shape:

{
  "topic": "...",
  "executive_summary": "...",
  "sections": [
    {
      "section_title": "...",
      "content": "...",
      "citations": ["S1", "S2"]
    }
  ],
  "sources": [
    {
      "id": "S1",
      "title": "...",
      "url": "..."
    }
  ]
}

Runtime Architecture

Core Components

• LangGraph: stateful orchestration and conditional routing.
• AgentState (TypedDict): shared state across workflow nodes.
• LangChain + OpenAI-compatible chat client: planner, analyzer, reporter calls.
• Pydantic v2: schema validation for planner input and final output.
• Google Custom Search: retrieval layer for web evidence.

Node Responsibilities

• planner: proposes next query, num_results, and optional freshness.
• search_tool: fetches and normalizes hits; deduplicates and appends context.
• analyzer: decides CONTINUE or FINISH; can adjust follow-up retrieval.
• reporter: composes the final structured report.

Data Contracts

Search Input

• query: str
• num_results: int (1..10)
• freshness: str | None

Final Output (FinalReport)

• topic: str
• executive_summary: str
• sections: List[ReportSection]
• sources: List[Source]

Provider Notes

Default: OpenRouter

• LLM_* variables are recommended.
• Legacy OPENROUTER_* variables are still supported for compatibility.

Optional Local Alternative: Venice

export LLM_BASE_URL=https://api.venice.ai/api/v1
export LLM_MODEL=venice:uncensored
export LLM_API_KEY=...

Troubleshooting

• 401 auth failure:
  • API key is invalid for the selected provider.
• 404 model unavailable:
  • Model ID is not available on that provider.
• 429 throttling:
  • Provider rate limit hit; retry or switch model/provider.
• OPENROUTER_MODEL='openrouter/free' is not a concrete model id:
  • Set an explicit model ID.
• Recursion limit errors:
  • Increase --max-iters or rerun (CLI applies a safe recursion multiplier).

What I Learned

• Prompt reliability requires deterministic fallbacks when models ignore strict schema output.
• Orchestration complexity grows quickly when loop control, stopping conditions, and retries interact.
• Structured validation is essential for turning model output into operationally useful JSON.
• External API constraints (auth, rate limits, model availability) must be handled explicitly at the edges.
• Retrieval and summarization quality are bottlenecks that need iterative tuning and evaluation.

Project Layout

• src/sra/cli.py: CLI and UI startup commands.
• src/sra/server.py: FastAPI UI backend and /api/run endpoint.
• src/sra/config.py: environment parsing and provider validation.
• src/sra/graph.py: LangGraph workflow and node logic.
• src/sra/tools.py: Google search integration and retry logic.
• src/sra/schemas.py: Pydantic data models.
• src/sra/state.py: shared graph state definition.