fintech-agent-observability-evaluation

KO | EN

gitlite — search

#javascript #ai #android #machine-learning #linux #docker #nodejs #windows #awesome #kubernetes #awesome-list #security #mcp #open-source

fintech-agent-observability-evaluation

(★ 12)

No description available.

fintech-agent-observability-evaluation 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Agent Observability, Evaluation & Safety Workshop
## Learn LangSmith · DeepEval · Guardrails AI · Presidio

A **4-hour hands-on workshop** teaching production-grade AI agent hardening through a concrete, runnable project: a multi-agent FinTech customer support system for SecureBank.

---

## What You'll Learn

| Concept | What It Is | How We Use It |
|---|---|---|
| **LangSmith** | LLM observability & evaluation platform | Trace every LLM call, debug failures, run A/B experiments |
| **DeepEval** | Open-source evaluation framework (pytest-native) | Faithfulness, hallucination, answer relevancy, G-Eval metrics |
| **Guardrails AI** | Output validation framework with 50+ validators | RegexMatch (SSN), ToxicLanguage, CompetitorCheck |
| **Presidio** | Microsoft's PII detection & redaction engine | Detect and redact SSNs, names, emails, credit cards |
| **MRR** | Mean Reciprocal Rank — retrieval quality metric | Measure how quickly the retriever surfaces the right document |
| **tiktoken** | OpenAI's token counting library | Count tokens locally, measure cost per query |

---

## The Scenario

**Company**: SecureBank (FinTech customer support)

A customer sends a question. The multi-agent system classifies intent, routes to the right specialist agent, and responds — all traced, evaluated, guarded, and cost-optimized.

| Agent | Role | Data Source | Example Query |
|---|---|---|---|
| **Supervisor** | Classifies intent → routes query | LLM classification | *(internal — runs on every query)* |
| **Policy Agent** | Answers policy questions via RAG | 4 policy documents (Chroma) | "What is the overdraft fee?" |
| **Account Agent** | Looks up account details | Mock database (3 accounts) | "What is the balance on ACC-12345?" |
| **Escalation Agent** | Empathetic handoff for complaints | LLM generation (no context) | "I'm furious! Nobody is helping!" |

### Mock Accounts

```
ACC-12345: Alice Johnson, Premium Checking, $12,450.75, active
ACC-67890: Bob Smith, Basic Checking, $234.50, active
ACC-11111: Carol Davis, High-Yield Savings, $85,320.00, FROZEN (fraud review)
```

---

## Project Structure

Modules are lettered in teaching order. Each module introduces new production hardening techniques applied to the same FinTech agent.

```
fintech-agent-observability-evaluation/
│
├── project/                               # THE AGENT — pre-built, shared by all modules
│   ├── fintech_support_agent.py           # Multi-agent system (supervisor + 3 agents)
│   └── documents/                         # Policy knowledge base (RAG source)
│       ├── account_fees.md                # Account types, overdraft fees, ATM fees
│       ├── loan_policy.md                 # Personal/auto loans, rates, eligibility
│       ├── fraud_policy.md                # Fraud detection, disputes, liability
│       └── transfer_policy.md             # Wire transfers, ACH, P2P, limits
│
├── module_a_observability/                # MODULE A — See what's happening
│   ├── README.md                          # Module guide for learners
│   ├── notes.md                           # Reference: observability, traces, monitoring
│   ├── demo.py                            # Silent failure demo + tagging runs
│   ├── exercise.py                        # Setup tracing, inspect traces, debug errors
│   └── solution.py                        # Complete solution
│
├── module_b_evaluation/                   # MODULE B — Measure quality
│   ├── README.md                          # Module guide for learners
│   ├── notes.md                           # Reference: datasets, evaluators, MRR, DeepEval, G-Eval
│   ├── eval_dataset.py                    # Shared evaluation datasets & helpers
│   ├── demo.py                            # Dataset creation + A/B experiments
│   ├── exercise.py                        # Evaluators, MRR, DeepEval metrics, G-Eval empathy
│   └── solution.py                        # Complete solution
│
├── module_c_guardrails/                   # MODULE C — Prevent bad outputs
│   ├── README.md                          # Module guide for learners
│   ├── notes.md                           # Reference: guardrails spectrum, Presidio, compliance
│   ├── demo.py                            # Before/after guardrail demo
│   ├── exercise.py                        # Guardrails AI validators + Presidio PII redaction
│   └── solution.py                        # Complete solution
│
├── module_d_cost_optimization/            # MODULE D — Optimize cost
│   ├── README.md                          # Module guide for learners
│   ├── notes.md                           # Reference: token economics, optimization patterns
│   ├── demo.py                            # Before/after cost comparison
│   ├── exercise.py                        # Token counting, cost measurement, comparison table
│   ├── solution.py                        # Complete solution
│   ├── audit_log.jsonl                    # Audit log output (generated by demo)
│   └── audit_log_solution.jsonl           # Audit log output (generated by solution)
│
├── .env.example                           # Copy to .env and add your API keys
└── requirements.txt                       # All Python dependencies
```

If module files feel overwhelming, start with the README inside each module folder.

### Notes live inside each module

Each module has a `notes.md` file with detailed reference documentation for that module's concepts.

| Module | Notes | Topics Covered |
|---|---|---|
| `module_a_observability/notes.md` | Observability deep dive | Traces, runs, run trees, debugging walkthrough, monitoring, sampling |
| `module_b_evaluation/notes.md` | Evaluation deep dive | Datasets, evaluators, MRR formula, DeepEval vs LangSmith, G-Eval criteria |
| `module_c_guardrails/notes.md` | Guardrails deep dive | Four strategies (Regex→Moderation→ML/NER→LLM), Guardrails AI, Presidio, GDPR/HIPAA patterns |
| `module_d_cost_optimization/notes.md` | Cost deep dive | Token economics, cost structure, before/after methodology, 4 optimization patterns |

---

## Quick Start

### 1. Prerequisites

- Python 3.12 (3.14 is **not** supported — `tiktoken` and other packages lack pre-built wheels)
- An OpenAI API key
- A free LangSmith account

**Verify installation:**
```bash
python --version  # should be 3.12.x
```

### 2. Clone or open this repo

```bash
git clone <your-repo-url>
cd fintech-agent-observability-evaluation
```

### 3. Create a virtual environment

> **Important**: Use Python 3.12. On Windows, if your default `python` points to a different version, use `py -3.12` instead.

**Windows (PowerShell):**
```powershell
py -3.12 -m venv .venv
```

**macOS/Linux:**
```bash
python3.12 -m venv .venv
```

### 4. Activate the virtual environment

**Windows (PowerShell):**
```powershell
Set-ExecutionPolicy -Scope Process -ExecutionPolicy Bypass; .\.venv\Scripts\Activate.ps1
```

**Windows (CMD):**
```bat
.venv\Scripts\activate.bat
```

**macOS/Linux:**
```bash
source .venv/bin/activate
```

> **Tip**: Your prompt will change to show `(.venv)` when the environment is active.

### 5. Install dependencies

```bash
python -m pip install --upgrade pip
pip install -r requirements.txt
```

### 6. Configure API keys

**Windows (PowerShell):**
```powershell
Copy-Item .env.example .env
```

**macOS/Linux:**
```bash
cp .env.example .env
```

Then edit `.env` and set:
```env
OPENAI_API_KEY=sk-your-key-here
LANGCHAIN_API_KEY=lsv2_pt_your-langsmith-key-here
LANGCHAIN_TRACING_V2=true
LANGCHAIN_PROJECT=fintech-support-agent
```

Get your LangSmith API key at [smith.langchain.com](https://smith.langchain.com) (free Developer plan — 1 seat, 5K traces/month).

### 7. Install Guardrails Hub validators (Module C)

```bash
pip install -r requirements.txt
python -m spacy download en_core_web_lg
```

Before installing hub validators, you need a free Guardrails Hub token:

1. Go to https://hub.guardrailsai.com/keys and sign up / log in
2. Copy your token
3. Run `guardrails configure` and paste the token when prompted

Then install the validators:

```bash
guardrails hub install hub://guardrails/regex_match
guardrails hub install hub://guardrails/toxic_language
guardrails hub install hub://guardrails/competitor_check
```

> **Note**: Presidio requires the spaCy `en_core_web_lg` model for NER-based PII detection. The `spacy download` command above installs it (~560 MB).

### 8. Run a smoke test

```bash
python module_a_observability/demo.py
```

If Module A runs and you see traces in LangSmith, your environment is ready.

### 9. Run the workshop modules in order

```bash
# MODULE A: Observability — trace the agent, debug silent failures
python module_a_observability/demo.py        # Instructor demo
python module_a_observability/exercise.py    # Your turn

# MODULE B: Evaluation — measure quality with LangSmith + DeepEval
python module_b_evaluation/demo.py           # Create dataset + A/B experiments
python module_b_evaluation/exercise.py       # Evaluators, MRR, DeepEval, G-Eval

# MODULE C: Guardrails — prevent PII leaks and unsafe outputs
python module_c_guardrails/demo.py           # Before/after demo
python module_c_guardrails/exercise.py       # Guardrails AI + Presidio

# MODULE D: Cost Optimization — measure and reduce token costs
python module_d_cost_optimization/demo.py    # Before/after cost comparison
python module_d_cost_optimization/exercise.py # Build your own comparison
```

---

## Architecture Deep Dive

### Multi-Agent Flow

```
python module_a_observability/demo.py
    │
    └── project/fintech_support_agent.py
          │
          ├── LangGraph: START → classify_intent → [policy | account | escalation] → END
          │
          ├── classify_intent (supervisor)
          │     ├── ChatOpenAI(model="gpt-4o-mini")
          │     ├── System prompt: "Classify into policy/account_status/escalation"
          │     └── Returns intent string → routes to specialist
          │
          ├── policy_agent (RAG)
          │     ├── Chroma retriever → top-k document chunks
          │     ├── format_docs → concatenate retrieved context
          │     └── ChatOpenAI → answer grounded in policy docs
          │
          ├── account_agent (DB lookup)
          │     ├── Regex extract ACC-XXXXX from query
          │     ├── MOCK_ACCOUNTS dict lookup
          │     └── ChatOpenAI → format account data as response
          │
          └── escalation_agent (empathetic handoff)
                ├── ChatOpenAI → empathetic response
                └── No context retrieval — just empathy + contact info
```

### Full Production Pipeline (After All Modules)

```
Customer Query
    │
    ▼
INPUT GUARDRAIL (Module C)
    ├── Presidio: redact PII from query
    └── Content safety: check for harmful input
    │
    ▼
MULTI-AGENT GRAPH (project/)     ──────►  LANGSMITH TRACING (Module A)
    ├── Supervisor: classify intent               │
    ├── Policy Agent: RAG generation              ├── Every run traced
    ├── Account Agent: DB lookup                  ├── Token counts captured
    └── Escalation Agent: handoff                 └── Latency measured
    │
    ▼
OUTPUT GUARDRAIL (Module C)
    ├── Guardrails AI: RegexMatch, ToxicLanguage, CompetitorCheck
    └── Presidio: redact any PII in response
    │
    ▼
Response to User

CONTINUOUS EVALUATION (Module B)          COST MONITORING (Module D)
    ├── Curate failing traces              ├── Token counting (tiktoken)
    ├── Evaluation datasets                ├── Before/after comparison
    ├── LangSmith evaluators               └── Projected savings
    ├── DeepEval (CI/CD)
    └── G-Eval (custom criteria)
```

### Cost Structure (Per Query)

```
POLICY QUERY: "What is the overdraft fee?"
  Supervisor:    ~100 prompt + ~3 completion tokens     ← cheap, runs every query
  Policy Agent:  ~800-1,500 prompt + ~50-150 completion ← expensive (RAG context)
  TOTAL:         ~$0.0003 (GPT-4o-mini)

ACCOUNT QUERY: "What is the balance on ACC-12345?"
  Supervisor:    ~100 prompt + ~3 completion tokens
  Account Agent: ~300 prompt + ~70 completion tokens
  TOTAL:         ~$0.0001 (GPT-4o-mini)

ESCALATION: "I'm furious! Nobody is helping!"
  Supervisor:       ~100 prompt + ~3 completion tokens
  Escalation Agent: ~120 prompt + ~100 completion tokens
  TOTAL:            ~$0.0001 (GPT-4o-mini)
```

---

## Workshop Schedule (4 Hours)

| Time | Module | Topic | Key Files |
|---|---|---|---|
| 0:00–0:10 | Intro | What we're building, setup check | `README.md`, `.env` |
| 0:10–1:15 | Module A | Agent observability with LangSmith | `module_a_observability/` |
| 1:15–1:20 | Break | | |
| 1:20–2:50 | Module B | Evaluation: datasets, MRR, DeepEval, G-Eval | `module_b_evaluation/` |
| 2:50–3:00 | Break | | |
| 3:00–3:50 | Module C | Input & output guardrails (4 strategies) + Presidio PII | `module_c_guardrails/` |
| 3:50–4:00 | Module D | Cost optimization (before/after) + wrap-up | `module_d_cost_optimization/` |

---

## Key Files Reference

| File | Key Function | What It Does |
|---|---|---|
| `project/fintech_support_agent.py` | `build_support_agent()` | Builds LangGraph multi-agent system (supervisor + 3 agents) |
| `project/fintech_support_agent.py` | `ask(app, query)` | Helper to invoke the agent with a query |
| `project/fintech_support_agent.py` | `MOCK_ACCOUNTS` | 3 mock bank accounts (ACC-12345, ACC-67890, ACC-11111) |
| `project/fintech_support_agent.py` | `SupportState` | TypedDict defining state flowing through the graph |
| `project/documents/*.md` | — | 4 policy documents embedded into Chroma for RAG |
| `module_b_evaluation/demo.py` | `evaluate()` | Creates LangSmith dataset and runs A/B experiments |
| `module_c_guardrails/solution.py` | `guarded_pipeline()` | Full pipeline: moderation → regex → injection → agent → Guardrails AI → Presidio |
| `module_d_cost_optimization/demo.py` | `run_measurement()` | Runs queries and captures token usage for before/after |

---

## Troubleshooting

### Dependency Installation Issues

#### `ResolutionImpossible` / Conflicting Dependencies

The pinned versions in `requirements.txt` are tested to work together. If you see dependency conflicts, ensure you are using the provided `requirements.txt` without modifications. Key version constraints to be aware of:

| Package | Constraint | Reason |
|---|---|---|
| `langsmith` | `<0.3,>=0.1.17` | `langchain` 0.3.x requires `langsmith<0.3` |
| `python-dotenv` | `>=1.1.1` | `deepeval>=3.8.9` requires `python-dotenv>=1.1.1` |
| `guardrails-ai` | Installed from GitHub `@v0.6.0` | PyPI is quarantined (CVE-2026-45758). v0.6.0 is compatible with langchain 0.3.x (`langchain-core<0.4`). |

> **Tip**: `langchain-core` and `langchain-text-splitters` should **not** be pinned directly — they are resolved automatically as transitive dependencies of `langchain` and `langchain-openai`.

If you still encounter conflicts, recreate your virtual environment from scratch:
```bash
# Remove and recreate the venv
rm -rf .venv               # macOS/Linux
Remove-Item -Recurse .venv # Windows PowerShell

python -m venv .venv
# Activate (see Step 4), then:
python -m pip install --upgrade pip
pip install -r requirements.txt
```

#### `error: Microsoft Visual C++ 14.0 or greater is required` (Windows)

The `chroma-hnswlib` package (required by `chromadb`) must be compiled from C++ source on Windows. This requires the **Microsoft Visual C++ Build Tools**.

**Fix — install build tools via `winget`:**
```powershell
winget install Microsoft.VisualStudio.2022.BuildTools --override "--passive --add Microsoft.VisualStudio.Workload.VCTools --add Microsoft.VisualStudio.Component.VC.Tools.x86.x64"
```

Or download manually from [https://visualstudio.microsoft.com/visual-cpp-build-tools/](https://visualstudio.microsoft.com/visual-cpp-build-tools/) and select the **"Desktop development with C++"** workload.

After installation, **close and reopen your terminal**, then re-run:
```powershell
pip install -r requirements.txt
```

#### `fatal error C1083: Cannot open include file: 'float.h'` (Windows)

If you already have Visual C++ Build Tools installed but the `chroma-hnswlib` build fails with:
```
fatal error C1083: Cannot open include file: 'float.h': No such file or directory
```

This means the **Windows 10/11 SDK** is missing. The MSVC compiler is installed but it cannot find the Universal CRT headers (`float.h`, `stdlib.h`, etc.) that ship with the Windows SDK.

**Fix — install the Windows SDK:**

*Option A — Visual Studio Installer (recommended):*
1. Open **Visual Studio Installer** (search for it in the Start menu)
2. Click **Modify** on **"Build Tools for Visual Studio 2022"**
3. Under **Individual Components**, check **"Windows 10 SDK"** (or **"Windows 11 SDK"**)
4. Click **Modify** and wait for the install to complete

*Option B — via `winget`:*
```powershell
winget install Microsoft.WindowsSDK.10.0.22621
```

After installing the SDK, **close and reopen your terminal**, then re-run:
```powershell
pip install -r requirements.txt
```

> **How to verify the fix**: After installing, check that the SDK headers exist:
> ```powershell
> Test-Path "C:\Program Files (x86)\Windows Kits\10\Include"
> ```
> This should return `True`.

#### C++ Compiler Errors on macOS / Linux

The same `chroma-hnswlib` compilation can fail on other platforms if a C++ compiler is missing:

**macOS** — install Xcode command-line tools:
```bash
xcode-select --install
```

**Linux (Debian/Ubuntu)** — install build essentials:
```bash
sudo apt-get update && sudo apt-get install -y build-essential
```

**Linux (Fedora/RHEL)**:
```bash
sudo dnf groupinstall "Development Tools"
```

After installing the compiler, re-run `pip install -r requirements.txt`.

---

### Common Runtime Errors

**`ModuleNotFoundError: No module named 'langchain'`**
```bash
pip install -r requirements.txt
```

**`AuthenticationError` from OpenAI**
```bash
# Check your .env file has the correct key
cat .env  # or: type .env (Windows)
```

**LangSmith traces not appearing**
```bash
# Ensure these are set in your .env:
LANGCHAIN_TRACING_V2=true
LANGCHAIN_API_KEY=lsv2_pt_your-key
```

**`ModuleNotFoundError: No module named 'deepeval'`**
```bash
pip install deepeval
```

**Guardrails Hub validators not found**
```bash
pip install -r requirements.txt  # installs guardrails-ai from GitHub (PyPI is quarantined)
guardrails configure  # paste your free token from https://hub.guardrailsai.com/keys
guardrails hub install hub://guardrails/regex_match
guardrails hub install hub://guardrails/toxic_language
guardrails hub install hub://guardrails/competitor_check
```

**Presidio not installed**
```bash
pip install presidio-analyzer presidio-anonymizer
```

**PowerShell `UnauthorizedAccess` error activating venv**
```powershell
Set-ExecutionPolicy -Scope Process -ExecutionPolicy Bypass; .\.venv\Scripts\Activate.ps1
```

**`FileNotFoundError` running module scripts**
```bash
# Run from the project root directory, not from inside a module folder
cd "fintech-agent-observability-evaluation"
python module_a_observability/demo.py
```

---

*Built for the AI Agent Systems Workshop — teaching production hardening through a FinTech multi-agent support system.*

fintech-agent-observability-evaluation

fintech-agent-observability-evaluation 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

fintech-agent-observability-evaluation

fintech-agent-observability-evaluation 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README