Agentic-AI-Pipeline

Agentic Multi‑Stage Bot - A Research & Outreach Agent 🤖

A full‑fledged, multi‑stage, agentic chatbot that plans → discovers → reasons → uses tools → learns. The reference task baked into this repo is a Research & Outreach Agent (“DossierOutreachAgent”): given a topic/company, it builds a compact, cited briefing and (optionally) drafts an outreach email, saving artifacts to disk.

Key Features
7‑Layer Architecture
Agent Graph (LangGraph)
Repository Layout
Quickstart
Configuration
Running
CLI Utilities
HTTP API
Tools & Capabilities
Memory & Feedback
Extending the System
Testing & Quality
GitHub Actions
Security & Ops
Troubleshooting
Roadmap Ideas

Key Features

Multi‑stage reasoning & planning with LangGraph state machine (plan → decide → act → tools → reflect → finalize).
Autonomous tool use & tool‑chaining (web search, URL fetch, calculator, file write, email draft, KB search/add).
Two memory systems:
- SQLite conversation store (turn history, feedback).
- ChromaDB vector KB for RAG and persistent knowledge.
Web experience layer: minimal FastAPI + SSE streaming UI (no framework lock‑in).
CLI for ingestion & demos.
Rate limiting (token bucket) per chat.
Batteries‑included tests (unit + smoke), lint, formatting, and ready Makefile targets.
Opinionated architecture: 7‑layer separation, single action per step, reflect‑first finalization.
Extensible: add new tools, nodes, or agent profiles without breaking the core.
Production ready: designed for real‑world use with security, observability, and extensibility in mind.
Open Source: Apache licensed, built with community contributions in mind.

7-Layer Architecture

This system is designed with a 7-layer architecture to separate concerns and allow for easy extension. Each layer has a specific role, and they communicate through well-defined interfaces. The core agent logic is implemented in LangGraph, which orchestrates the reasoning and tool use.

[!NOTE] Agent ≠ LLM. This system implements a layered design that’s easy to extend and operate.

Experience (UI/API)
   └─ FastAPI + SSE → web UI, CLI, REST
Discovery
   └─ WebSearch, WebFetch, KB search (RAG)
Agent Composition
   └─ Agent profile (persona, objective, capabilities)
Reasoning & Planning
   └─ LangGraph plan→decide→act→reflect→finalize
Tool & API
   └─ Calculator, FileWrite, Emailer, Web tools, KB add/search
Memory & Feedback
   └─ SQLite (turns, feedback), ChromaDB (vector KB)
Infrastructure
   └─ Logging, rate limiting, config, packaging

Directory mapping

Layer	Where
Experience	`src/agentic_ai/app.py`, `web/`
Discovery	`src/agentic_ai/tools/webtools.py`, `tools/knowledge.py`
Agent Composition	`src/agentic_ai/layers/composition.py`
Reasoning & Planning	`src/agentic_ai/layers/reasoning.py`, `graph.py`
Tools & API	`src/agentic_ai/tools/*`, `layers/tools.py`
Memory & Feedback	`src/agentic_ai/memory/*`, `layers/memory.py`
Infrastructure	`src/agentic_ai/infra/*`, `config.py`, `pyproject.toml`

By separating these layers, we ensure that each component can evolve independently while maintaining a clear contract with the others. This modularity allows for easy testing, debugging, and extension of the system.

7-Layer Architecture

Credit: Anil Inamdar for the original 7-layer architecture diagram and for the inspiration behind the agentic design principles.

Agent Graph (LangGraph)

The agent runs a typed state machine driving the LLM and tools. It uses LangGraph to orchestrate the reasoning and tool use, with a focus on single atomic actions per step. This design ensures clarity in execution and allows for easy debugging and extensibility.

Agentic AI Graph

Dynamic Mermaid graph of the agent’s architecture:

flowchart LR

%% =========================
%% EXPERIENCE LAYER
%% =========================
subgraph EXP[Experience Layer FastAPI SSE]
direction LR
IN[User prompt]
RL{Rate limit check}
STREAM[Token stream SSE]
FB[Feedback API]
IN --> RL
STREAM --> FB
end

%% =========================
%% MEMORY LAYER
%% =========================
subgraph MEM[Memory and Feedback]
direction TB
SQL[(SQLite messages)]
VEC[(Chroma vector KB)]
CIT[Citations store]
end

RL -->|ok| START
RL -->|429| HALT[Backoff and retry]

%% =========================
%% REASONING AND PLANNING
%% =========================
subgraph GRAPH[Reasoning and Planning LangGraph]
direction LR
START((START))
PLAN[PLAN create plan]
DECIDE{DECIDE next action}
ACT[ACT bind tool emit call]
TOOLS((ToolNode executes call))
REFLECT[REFLECT briefing or next]
FINALIZE([FINALIZE result])

START --> PLAN
PLAN --> DECIDE

DECIDE -->|search| ACT
DECIDE -->|fetch| ACT
DECIDE -->|kb_search| ACT
DECIDE -->|calculate| ACT
DECIDE -->|write_file| ACT
DECIDE -->|draft_email| ACT
DECIDE -->|finalize| FINALIZE
DECIDE -->|otherwise| REFLECT

ACT --> TOOLS
TOOLS --> REFLECT

REFLECT -->|done| FINALIZE
REFLECT -->|not done| DECIDE

PLAN -.-> STREAM
DECIDE -.-> STREAM
REFLECT -.-> STREAM
FINALIZE -.-> STREAM
end

PLAN -. use kb hits .-> VEC

%% =========================
%% TOOL AND API LAYER
%% =========================
subgraph TOOLS_LAYER[Tool and API Layer]
direction TB
TSEARCH[web search]
TFETCH[web fetch]
TKBS[kb search]
TKBADD[kb add]
TCALC[calculator]
TWRITE[file write]
TEMAIL[emailer]
end

TOOLS -->|query| TSEARCH
TOOLS -->|url| TFETCH
TOOLS -->|semantic| TKBS
TOOLS -->|ingest| TKBADD
TOOLS -->|expr| TCALC
TOOLS -->|path content| TWRITE
TOOLS -->|to subject body| TEMAIL

TKBS -->|results| PLAN
TKBADD --> VEC
TSEARCH --> TFETCH
TFETCH -->|text| PLAN
TFETCH --> CIT
TWRITE -->|artifact path| PLAN
TEMAIL -->|eml path| PLAN

IN -->|save user| SQL
FINALIZE -->|save assistant| SQL
FB -->|save feedback| SQL

The graph consists of the following key nodes:

PLAN: creates a short action plan (3–6 steps), with KB context pre‑retrieval.
DECIDE: selects the next atomic action (token): search, fetch, kb_search, calculate, write_file, draft_email, or finalize.
ACT: binds LLM to tool schema and forces a single precise tool call.
TOOLS: concrete execution via LangGraph ToolNode.
REFLECT: either emits BRIEFING: ... (final) or NEXT: <action>.
FINALIZE: completes, persisting the result.

This design ensures that each step is clear and focused, allowing for easy debugging and understanding of the agent’s behavior. The use of LangGraph allows for powerful state management and tool orchestration, making it a robust solution for building agentic systems.

Repository Layout

src/
  agentic_ai/
    app.py                # FastAPI app (Experience)
    graph.py              # LangGraph runner
    cli.py                # CLI utilities (ingest/demo)
    config.py             # Pydantic settings
    infra/
      logging.py          # Rotating logs
      rate_limit.py       # Token bucket
    memory/
      sql_store.py        # SQLite conversation store
      vector_store.py     # ChromaDB KB
    layers/
      composition.py      # Agent profile
      reasoning.py        # Plan→Decide→Act→Reflect graph
      memory.py           # Memory orchestrator helpers
      tools.py            # Tool registry
    tools/
      webtools.py         # DuckDuckGo search + URL fetch & extract
      ops.py              # Calculator, FileWrite, Emailer (mock)
      knowledge.py        # KB add/search tools
web/
  index.html styles.css app.js
data/
  seed/                   # Non-empty RAG seeds
  agent_output/          # FileWrite artifacts
  emails/                # Emailer artifacts
tests/
  ... (unit & smoke)

[!IMPORTANT] This layout separates concerns clearly, with src/agentic_ai containing the core logic, web/ for the UI, and data/ for input/output artifacts. The tests/ directory holds unit and smoke tests to ensure code quality.

Quickstart

To quickly get started with the Agentic AI bot, follow these steps:

# 1) Create venv & install
python -m venv .venv
. .venv/bin/activate
pip install -U pip
pip install -r requirements.txt

# 2) Configure model keys
cp .env.example .env
# Edit .env with your OPENAI_API_KEY (or ANTHROPIC_API_KEY) and preferences

# 3) Ingest seed knowledge into vector store
make ingest

# 4) Run the server
make run
# Open http://localhost:8000 and ask:
# "Build a competitive briefing on ACME Robotics and draft a short outreach email."

CLI demo (streams to stdout)

. .venv/bin/activate
python -m agentic_ai.cli demo "Summarize top AMR vendors for grocery DCs with citations"

This will start the agent, which will plan, search, and reflect on the topic, providing a compact briefing with citations. You can also use the CLI to ingest additional documents into the vector knowledge base.

Configuration

The app uses Pydantic Settings and .env.

MODEL_PROVIDER=openai          # openai | anthropic
OPENAI_API_KEY=sk-...          # required if MODEL_PROVIDER=openai
OPENAI_MODEL_CHAT=gpt-4o-mini
OPENAI_MODEL_EMBED=text-embedding-3-small

ANTHROPIC_API_KEY=sk-ant-...   # required if MODEL_PROVIDER=anthropic
ANTHROPIC_MODEL_CHAT=claude-3-5-sonnet-latest

CHROMA_DIR=.chroma
SQLITE_PATH=.sqlite/agent.db

APP_HOST=0.0.0.0
APP_PORT=8000

You can switch providers at runtime by changing MODEL_PROVIDER and the corresponding keys/models.

Running

We recommend using a virtual environment to isolate dependencies. The provided Makefile simplifies setup and common tasks:

Makefile targets

make setup – create venv, install deps, copy .env.example
make ingest – add files under data/seed/ to the vector KB
make run – launch FastAPI on configured host/port
make test – run unit tests
make format / make lint – apply/check ruff (imports + formatting)

Run the server

. .venv/bin/activate
uvicorn src.agentic_ai.app:app --host

This will start the FastAPI server on http://localhost:8000, where you can interact with the agent via the web UI or API.

CLI Utilities

This repository also provides a CLI for quick interactions and ingestion of knowledge:

# Ingest any folder recursively (adds text up to 10k chars per file)
python -m agentic_ai.cli ingest ./data/seed

# Run a single prompt through the graph (streams responses)
python -m agentic_ai.cli demo "Give me a briefing on collaborative robots with citations"

The CLI allows you to quickly test the agent’s capabilities without needing to set up the web UI. It streams responses directly to the console, making it easy to see how the agent processes prompts and uses tools.

HTTP API

A simple HTTP API is provided for programmatic access to the agent’s capabilities. It supports both synchronous and streaming interactions:

`GET /api/new_chat`

Returns a new chat_id.

`POST /api/chat` (SSE stream)

Request body

{ "chat_id": "uuid-string", "message": "Your prompt" }

Stream events

event: token – incremental text chunks
event: done – {"chat_id": "..."}

`POST /api/ingest`

Add an internal KB document.

{ "id": "doc-123", "text": "content...", "metadata": { "source": "..." } }

`POST /api/feedback`

Attach a rating/comment to a chat/message.

{ "chat_id": "uuid", "message_id": null, "rating": 1, "comment": "👍" }

This API allows you to integrate the agent into your applications, enabling programmatic access to its reasoning and tool capabilities. The SSE stream provides real-time updates, making it suitable for interactive applications. Feedback can be used to improve the agent’s performance over time.

Tools & Capabilities

The agent uses a set of atomic tools to perform actions. Each tool is designed for a specific purpose, and they can be chained together to achieve complex tasks. The tools are registered in the tools.py layer and can be invoked by the agent during reasoning.

Tool	Name	Purpose	Input	Output
`web_search`	WebSearch	General discovery	NL query	JSON list of results
`web_fetch`	WebFetch	Extract readable text from URL	URL	Clean text
`kb_search`	KbSearch	Vector KB semantic search	NL query	JSON list of {id,text,metadata}
`kb_add`	KbAdd	Add doc to KB	JSON `{id,text,metadata}`	`"ok"`
`calculator`	Calculator	Safe math via `math` module	Expression	Result string
`file_write`	FileWrite	Persist artifacts under `data/agent_output/`	JSON `{path,content}`	Absolute path
`emailer`	Emailer (mock)	Draft/queue email as `.eml`	JSON `{to,subject,body}`	`.eml` path

Tool‑chaining is managed by the graph: DECIDE picks one atomic action per step, ACT binds tools and forces one structured call, TOOLS executes, REFLECT synthesizes or continues.

This design ensures that each tool is used in a controlled manner, allowing for clear reasoning and traceability of actions. The agent can also reflect on its actions and decide whether to finalize or continue with further steps.

Memory & Feedback

The agent has two memory systems to persist conversation history and knowledge:

Conversation turns: persisted in SQLite via SQLStore (messages table holds role, content, tool_call).
Vector KB: ChromaDB persistent collection; kb_add and kb_search power RAG and context for planning.
Feedback: optional feedback table (rating + comment), useful for reinforcement signals and post‑hoc analysis.

DB schema (simplified)

CREATE TABLE IF NOT EXISTS chats (
  id TEXT PRIMARY KEY,
  created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
  title TEXT
);
CREATE TABLE IF NOT EXISTS messages (
  id INTEGER PRIMARY KEY AUTOINCREMENT,
  chat_id TEXT,
  role TEXT,
  content TEXT,
  tool_call TEXT,
  created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP
);
CREATE TABLE IF NOT EXISTS feedback (
  id INTEGER PRIMARY KEY AUTOINCREMENT,
  chat_id TEXT,
  message_id INTEGER,
  rating INTEGER,
  comment TEXT,
  created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP
);

Feel free to extend the memory layer with additional tables or fields as needed. The agent can use this memory to maintain context across interactions, allowing for more coherent and informed responses. More details on the memory layer can be found in src/agentic_ai/memory/.

Extending the System

To extend the system, you can add new tools, nodes, or modify the agent profile. Here are some guidelines:

Add a new Tool

Create src/agentic_ai/tools/<your_tool>.py implementing a BaseTool with _run.
Register it in src/agentic_ai/layers/tools.py::registry().
Teach the decision policy by adding the token to DECIDE’s prompt and mapping in ACT.

Add a new Node (LangGraph)

Add a function (state) -> state, g.add_node("name", fn), and wire edges.
Prefer small atomic nodes; use reflect to decide if you have enough evidence to finalize.

Change the Agent Profile

Edit layers/composition.py::PROFILE (persona, objective, capabilities).
Provide role or safety constraints in SYSTEM prompt in layers/reasoning.py.

Use Another Vector DB / Embeddings

Replace VectorStore implementation and/or embedding function.
Adjust ingestion (cli.py) and KB tool wiring accordingly.

The system is designed to be modular and extensible, allowing you to adapt it to your specific use cases. The LangGraph framework provides a powerful way to manage the agent’s reasoning and tool use, making it easy to add new capabilities without disrupting existing functionality.

[!TIP] Feel free to experiment with different tools, nodes, and configurations to tailor the agent to your needs. The modular design allows for easy testing and iteration, making it a flexible platform for building agentic systems!

Testing & Quality

This repository includes a comprehensive suite of tests to ensure code quality and functionality. The tests cover unit tests for individual components, as well as smoke tests for end-to-end scenarios.

Run all tests

. .venv/bin/activate
pytest -q

What’s covered

Memory round‑trips (SQLite + ChromaDB).
Tool smoke tests (calculator/file/email/KB).
API endpoint sanity (new_chat/ingest/feedback).
Registry integrity.
Graph reasoning with a Fake LLM (if available in langchain_core) to avoid API keys during CI.

Lint & Format

make format
make lint

GitHub Actions

This repository uses GitHub Actions for continuous integration (CI) to ensure code quality and functionality. The CI pipeline includes:

Linting: checks code style and formatting using ruff.
Testing: runs unit tests and smoke tests to verify functionality.
Type checking: ensures type safety using mypy.
Dependency management: checks for outdated dependencies and security vulnerabilities.
Documentation generation: builds the documentation from the code comments and README.
Code coverage: reports code coverage to ensure sufficient test coverage.
Release management: automatically creates releases based on tags.
Docker build: builds a Docker image for the application.
Security checks: runs security scans on the codebase.
and many more!

You can view the CI pipeline in the .github/workflows/ directory. This should provide you with a solid foundation for maintaining code quality and ensuring that your changes do not break existing functionality, and feel free to extend the CI pipeline with additional checks or steps as needed.

Security & Ops

We also incorporate DevSecOps best practices to ensure the system is secure and production-ready:

Rate limiting: simple token bucket (infra/rate_limit.py) to mitigate spam.
Logging: rotating file + console (.logs/agentic-ai.log).
Secrets: .env (never commit) via Pydantic Settings.
Isolation: venv recommended; for servers, run behind a reverse proxy (nginx/Caddy).
Outbound tools: web_fetch honors redirects and has timeouts; consider adding domain allowlists for production.

[!CAUTION] For production, consider:

API auth (JWT/OAuth) and per‑tenant namespaces for vector DB.

Audit logging of tool calls.

Observability (OpenTelemetry) and tracing via LangSmith or vendor of choice.

Dockerization and CI/CD (GitHub Actions) – the Makefile is a good start.

Rate limits on LLM calls to prevent abuse.

Bounded tool I/O: ensure tools have strict input/output limits to prevent abuse.

Security headers: set appropriate HTTP security headers (CSP, CORS, etc.) in FastAPI.

HTTPS: always run behind HTTPS in production (use nginx/Caddy).

Environment isolation: use separate environments for development, staging, and production.

Monitoring: set up monitoring for performance, errors, and usage patterns.

Backup: regularly back up the SQLite database and ChromaDB vector store.

Data retention: implement data retention policies for conversation history and feedback.

Compliance: ensure compliance with data protection regulations (GDPR, CCPA, etc.) if handling personal data.

Security reviews: conduct regular security reviews and vulnerability assessments.

Dependency management: keep dependencies up to date and monitor for vulnerabilities (e.g., using pip-audit).

Incident response: have an incident response plan in place for security breaches or data leaks.

This repository is designed to be a solid foundation for building agentic systems, with security and operational best practices in mind. However, always review and adapt the security measures based on your specific deployment environment and requirements. However, in production, make sure your agent is running behind a reverse proxy (like nginx or Caddy) to handle HTTPS, rate limiting, and security headers, and comply with standards like GDPR or CCPA if handling personal data.

Troubleshooting

If you encounter issues while running the agent, here are some common troubleshooting steps:

Model errors: ensure .env has the correct provider and API key.
Vector store empty? Run make ingest or POST /api/ingest.
Unicode or huge files on ingest: the CLI truncates to ~10k chars per file; tweak in cli.py.
Streaming stalls: check browser console and server logs; SSE requires no buffering proxies (disable proxy buffering or set proxy_buffering off in nginx).
Tool misuse: update DECIDE prompt and ACT mapping to strongly bias the correct tool.

If you still have issues, feel free to open an issue on GitHub with details about your setup and the error messages you’re encountering. The community is here to help!

Roadmap Ideas

We still have a lot of ideas for future improvements and features. Here are some potential roadmap items:

Multi‑agent composition: add specialist agents (e.g., “Citer”, “Editor”) and coordinate via LangGraph subgraphs.
Structured outputs: Pydantic validation for deliverables (JSON schemas for dossiers/emails).
Retrieval upgrades: hybrid search (BM25 + embeddings), reranking, summary chunking.
Learning loop: periodic fine‑tuning signals from feedback table or prompt cache.
Real email: integrate SMTP or provider SDK behind a feature flag.
Auth & tenancy: API keys per user, KB namespaces per tenant.

If you have ideas or contributions, feel free to open an issue or PR! We welcome community input to make this agentic system even better.

Credits

This project leverages several powerful libraries and frameworks to provide a robust agentic system:

LangGraph for robust agent graphs.
LangChain for tool abstractions.
FastAPI/Uvicorn for a sweet, fast developer experience.
ChromaDB/SQLite for pragmatic persistence.
DuckDuckGo Search for frictionless discovery.
OpenAI/Anthropic for chat LLMs.
and many more!

What makes this opinionated?

This system is designed with a few key principles in mind:

Single‑action per step: encourages deterministic tool traces and readable logs.
Reflect‑first finalization: no wall‑of‑text dumps; concise, cited briefings.
7‑layer separation: swap or extend layers without breaking others.
Extensible tools: add new tools or nodes without disrupting the core logic.
Opinionated defaults: preconfigured agent profile, tools, and memory systems to get you started quickly.
Production ready: designed for real‑world use with security, observability, and extensibility in mind.
Open Source: Apache licensed, built with community contributions in mind.
Community driven: we welcome contributions and ideas to improve the system.
Batteries included: unit tests, linting, formatting, and a Makefile to simplify common tasks.
Minimal web UI: no framework lock‑in, just FastAPI + SSE for streaming responses.
CLI utilities: for quick interactions and ingestion of knowledge.
Rate limiting: simple token bucket to prevent abuse.
Security best practices: logging, secrets management, and outbound tool safety.

This opinionated design allows for a clear and consistent development experience, making it easier to build and maintain agentic systems. The modular architecture ensures that you can adapt the system to your specific needs while still benefiting from the core features and best practices.

License

This project is licensed under the Apache License. You are free to use, modify, and distribute this code as long as you include the original license and copyright notice.

[!NOTE] Regardless of use, make sure you credit the original authors and contributors when using this code in your projects. Contributions are welcome, and we encourage you to share your improvements or extensions to the system.

Conclusion

This repository provides a solid foundation for building agentic systems with multi‑stage reasoning, autonomous tool use, and a focus on security and extensibility. The modular architecture allows for easy extension and customization, making it suitable for a wide range of applications.

We hope this project serves as a useful resource for your agentic AI projects. Whether you’re building research agents, outreach bots, or any other type of agentic system, the principles and patterns demonstrated here can help you get started quickly and effectively.

If you have any questions, suggestions, or contributions, please feel free to open an issue or pull request on GitHub. We welcome community input and are excited to see how you use this system in your own projects!

Happy building! We hope this repository serves as a useful resource for your agentic AI projects. If you have any questions or suggestions, feel free to reach out or contribute!

This site is open source. Improve this page.

Agentic-AI-Pipeline

Agentic Multi‑Stage Bot - A Research & Outreach Agent 🤖

Table of Contents

Key Features

7-Layer Architecture

Agent Graph (LangGraph)

Repository Layout

Quickstart

Configuration

Running

CLI Utilities

HTTP API

GET /api/new_chat

POST /api/chat (SSE stream)

POST /api/ingest

POST /api/feedback

Tools & Capabilities

Memory & Feedback

Extending the System

Add a new Tool

Add a new Node (LangGraph)

Change the Agent Profile

Use Another Vector DB / Embeddings

Testing & Quality

GitHub Actions

Security & Ops

Troubleshooting

Roadmap Ideas

Credits

What makes this opinionated?

License

Conclusion

`GET /api/new_chat`

`POST /api/chat` (SSE stream)

`POST /api/ingest`

`POST /api/feedback`