LLM-RAG Deep Agent using LangChain, LangGraph, LangSmith

Python LLM-RAG deep agent using LangChain, LangGraph and LangSmith built on Quart web microframework and served using Hypercorn ASGI and WSGI web server.

Infrastructure components:

All of the following components run on k8s cluster:

PostgreSQL for checkpoints and vector DB
(Optional) Chroma for vector DB
Neo4J for graph query
Ollama as LLM model server

Sources of information for RAG

Unstructured

Online blog posts ingested into vector database
Text strings extracted from SQL database ingested into Neo4J vector database
Google ground search

Structured

SQL database relationships ingested into Neo4J graph database

Environment

Add a .env with the following environment variables:

ENVIRONMENT=development
DB_USERNAME=
DB_PASSWORD=
NEO4J_AUTH=username/password
CHROMA_TOKEN=
LANGSMITH_TRACING="true"
LANGSMITH_API_KEY=
LANGSMITH_ENDPOINT="https://api.smith.langchain.com"
LANGSMITH_PROJECT=
OPENAI_API_KEY=
GOOGLE_CLOUD_PROJECT=
GOOGLE_CLOUD_LOCATION="asia-southeast1"
GOOGLE_API_KEY=
GEMINI_API_KEY=
VERTEX_API_KEY=
USER_AGENT="USER_AGENT"

Install tkinter:

$ sudo apt install -y python3.13-tk

Data Source Preparation:

PostgreSQL Database

Firstly, create an empty database "Langchain" in PostgreSQL

Database Migration

Copy env.py to migrations/ folder.
Set the values DB_foo in /etc/ragagent_config.json

run migrations initialization with db init command:

$ uv run alembic init migrations
$ cp env.py migrations
$ uv run alembic revision --autogenerate -m "Initial migration"
$ uv run alembic upgrade head

There will be 1 table, "users" created in the PostgreSQL database "Langchain" after the upgrade.

PostgresSQL PGVector

Run the following CLI command to crawl and ingest blogs into PostgreSQL PGVector.

$ uv run python -m src.rag_agent.RAGAgent -l

Neo4J Graph DB

Import data from CSV files into the database
Copy the files / folder into the pod /var/lib/neo4j/import
LoadNeo4J.sh will load data/Healthcare into neo4j-0 pod

Local Model

https://python.langchain.com/docs/how_to/local_llms/

Ollama

https://github.com/ollama/ollama
Download and install the app:

$ curl -fsSL https://ollama.com/install.sh | sh

Run the app with a model:

$ ollama run gpt-oss

Google VertexAI / GenAI

Install Google Cloud CLI:

$ pip3 install --upgrade google-cloud-aiplatform
$ sudo snap install google-cloud-cli --classic

Setup Google Cloud Authentication and Project:

$ gcloud init
$ gcloud auth application-default login
$ gcloud auth application-default set-quota-project <ProjectID>

Start the application:

To run the agent from console:

$ uv run python -m src.rag_agent.RAGAgent -h
usage: RAGAgent.py [-h] [-l] [-t] [-d] [-a] [-m] [-n] [-u] [-v] [-b] [-w]

LLM-RAG deep agent answering user questions about healthcare system and AI/ML

options:
  -h, --help            show this help message and exit
  -l, --load-urls       Load documents from URLs
  -t, --show-threads    Show history of all threads
  -d, --delete-threads  Delete history of all threads
  -a, --ai-ml           Ask questions regarding AI/ML which should be answered based on Lilian's blog
  -m, --mlflow          Ask questions regarding MLFlow
  -n, --neo4j-graph     Ask question with answer in Neo4J graph database store
  -u, --stream-updates  Stream updates instead of theh complete message
  -v, --neo4j-vector    Ask question with answers in Neo4J vector store
  -b, --neo4j           Ask question with answers in both Neo4J vector and graph stores
  -w, --wait-time       Ask hospital waiting time using answer from mock API endpoint

To run the ASGI/WSGI web application:

$ ./hypercorn.sh

HTTP/3 curl:

To build your own HTTP/3 curl: https://curl.se/docs/http3.html
Add Host header which is defined as server_names in hypercorn.toml

curl --http3-only --insecure -v https://localhost:4433/<path> -H "Host: khteh.com"

curl --http3-only --insecure -vv https://localhost:4433/chat/invoke -F 'prompt=:"What is task decomposition?"' -F 'file=@~/data/1.jpg' -F 'receipt=true'

TLS certificates

If run locally, ./hypercorn.sh would have generated the TLS cert into /tmp folder
If trying to access the application running in k8s cluster, copy the cert out of the pod into /tmp folder.

Chrome browser

Close ALL chrome browser (both tabs and windows)
Generate TLS certificate fingerprint:

$ fingerprint=`openssl x509 -pubkey -noout -in /tmp/server.crt |
        openssl rsa -pubin -outform der |
        openssl dgst -sha256 -binary | base64`

Start Chrome browser with QUIC protocol for HTTP/3:
$1 is URL. For example: https://localhost:4433

$  /opt/google/chrome/chrome --disable-setuid-sandbox --enable-quic --ignore-certificate-errors-spki-list=$fingerprint --origin-to-force-quic-on=${1#*//} $1

Home controller endpoints:

$ c3 -v https://localhost:4433/invoke -m 300 -X POST -d '{"message": "What is task decomposition?"}'

Hospital controller endpoints:

$ c3 -v https://localhost:4433/healthcare/invoke -m 300 -X POST -d '{"message": "Which hospital has the shortest wait time?"}'

Launch LangGraph Server

Checkpointed RAG

Configure langgraph.json with:

    "graphs": {
        "rag_agent": "./src/rag_agent/CheckpointedRAG.py:make_graph"
    },

RAG ReAct Agent

Configure langgraph.json with:

    "graphs": {
        "rag_agent": "./src/rag_agent/RAGAgent.py:make_graph"
    },

Run local Langgraph server

uv run python -m langgraph_cli dev

Main Agent:

Question & Answer RAG Deep Agent

Sub-Agent

RAG Deep Agent answers question from Vector and Graph Database

Answering questions from mocked-up API call:

output/user_questions.md:

Which hospital has the shortest wait time?

output/final_answer.md:

## Shortest Wait Time Hospital

| Hospital | Location | Wait Time |
|----------|----------|-----------|
| **Walton LLC** | Florida (FL) | **6 minutes** |

**Sources**

- Shortest wait time: *get_most_available_hospital* returned “Walton LLC” with a wait time of 6 minutes.
- Location: *HealthcareCypher* confirmed that Walton LLC is located in Florida.

Answering questions from Postgres Vector Store:

output/user_request_{timestamp}.md:

What is task decomposition?

---
## Task Decomposition

Task decomposition is the process of breaking a complex, high‑level task into a hierarchy of smaller, more manageable sub‑tasks or sub‑goals. In AI planning it is often formalised as hierarchical task network (HTN) planning, where a task is recursively refined until only primitive actions that can be directly executed remain [1].

### Purpose
1. **Manage complexity** – Large problems become tractable when expressed as a set of simpler components.\
2. **Enable hierarchical reasoning** – Allows planners or models to operate at different abstraction levels and reuse sub‑task solutions across domains.\
3. **Facilitate parallelism and scheduling** – Independent sub‑tasks can be dispatched to different agents or processors.\
4. **Improve interpretability** – A decomposition tree provides a clear rationale for each step taken by an autonomous system.

### Typical Methods

| Method | Main Idea | Representative Source |
|--------|-----------|------------------------|
| **Hierarchical Task Network (HTN) Planning** | Uses predefined *methods* to decompose a non‑primitive task into subtasks until only primitive actions remain (PDDL‑style). | Nau et al., *Automated Planning: Theory & Practice* [1] |
| **Chain‑of‑Thought (CoT) Prompting** | Large language models are prompted to “think step‑by‑step”, implicitly decomposing a problem into a sequential chain of reasoning steps. | Wei et al., *Chain‑of‑Thought Prompting* [2] |
| **Tree‑of‑Thoughts (ToT)** | Extends CoT by generating multiple possible reasoning steps per level, forming a tree that is searched (BFS/DFS) for the best solution. | Yao et al., *Tree of Thoughts* [3] |
| **LLM + Classical Planner (LLM+P)** | The LLM translates a natural‑language task into a formal planning language (e.g., PDDL); an external planner solves the decomposed problem, and the plan is translated back for execution. | Liu et al., *LLM‑Planner* [4] |
| **Hierarchical Reinforcement Learning (HRL)** | Learns policies at different levels of abstraction; a high‑level policy selects sub‑goals for a low‑level controller. | Sutton et al., *Between MDPs and Semi‑MDPs* [5] |
| **Program Synthesis / Sub‑task Generation** | An LLM generates code or scripts that implement sub‑tasks, which are then executed sequentially or in parallel. | Liu et al., *Program‑Aided Language Models* [6] |

### How the Methods Relate
- **Classical AI** (HTN, LLM+P) provides formal, verifiable decompositions using planning languages.\
- **Neural approaches** (CoT, ToT, HRL, program synthesis) infer useful sub‑tasks from data‑driven models, often without explicit symbolic representations.\
- **Hybrid systems** combine symbolic planners with LLMs, leveraging formal guarantees together with flexible, knowledge‑rich reasoning.

### Concise Summary
Task decomposition is a foundational AI technique that splits a complex objective into a hierarchy of simpler sub‑goals, facilitating planning, execution, and interpretability. Classical methods such as HTN planning use predefined methods and formal languages to guarantee soundness. Recent neural techniques—Chain‑of‑Thought, Tree‑of‑Thoughts, hierarchical reinforcement learning, and LLM‑assisted planners—derive decompositions from large language models, either as sequential reasoning steps or as branching search trees. Hybrid approaches integrate symbolic planners with LLMs to combine formal guarantees with flexible knowledge, enabling autonomous agents to tackle long‑horizon, multi‑step problems across robotics, NLP, and software synthesis.

---
### Sources
[1] Nau, D., et al., *Automated Planning: Theory & Practice – Hierarchical Task Networks*, 2003. https://doi.org/10.1017/CBO9780511807400
[2] Wei, X., et al., *Chain‑of‑Thought Prompting Elicits Reasoning in Large Language Models*, 2022. https://arxiv.org/abs/2201.11903
[3] Yao, S., et al., *Tree of Thoughts: Deliberate Problem Solving with Large Language Models*, 2023. https://arxiv.org/abs/2305.10601
[4] Liu, H., et al., *LLM‑Planner: Integrating Large Language Models with Classical Planners for Long‑Horizon Tasks*, 2023. https://arxiv.org/abs/2305.12923
[5] Sutton, R., et al., *Between MDPs and Semi‑MDPs: A Framework for Hierarchical Reinforcement Learning*, 1999. https://doi.org/10.1109/ICML.1999.748044
[5] Liu, P., et al., *Program‑Aided Language Models (PAL)*, 2022. https://arxiv.org/abs/2203.07814

output/user_request_{timestamp}.md:

What is MLFlow?

---
## What is MLflow?

MLflow is an open‑source platform that helps data‑science and engineering teams manage the **entire machine‑learning lifecycle** – from experimentation to reproducible runs, model versioning, and deployment. It is vendor‑neutral, runs anywhere (local machine, on‑premise clusters, cloud), and provides a unified set of APIs and UI components [1].

### Core Components

| Component | Purpose | Key Features |
|-----------|---------|--------------|
| **MLflow Tracking** | Logs experiments (parameters, metrics, artifacts, code version) and visualizes results through a UI or REST API [2]. | Python, R, Java APIs; auto‑logging; remote tracking server; searchable runs. |
| **MLflow Projects** | Packages a reproducible code bundle (environment + entry point) so anyone can run the same experiment on any platform [3]. | Uses Conda or Docker environments; `mlflow run` CLI; supports Git‑based projects. |
| **MLflow Models** | Standardizes model serialization and provides a common “flavor” interface to load models in Python, R, Java, or as a REST endpoint [4]. | Supports Scikit‑learn, TensorFlow, PyTorch, SparkML, H2O, etc.; model signatures; artifact storage. |
| **MLflow Model Registry** | Central hub for model versioning, stage transitions (Staging → Production → Archived), and lineage tracking [5]. | UI & API for approvals, commenting, and access control; integrates with CI/CD pipelines. |

### Typical Use Cases

1. **Experiment Tracking** – Compare many hyper‑parameter runs, visualise metrics, and share results across a team.
2. **Reproducible Pipelines** – Encode a training script, its dependencies, and data paths in an MLflow Project; anyone can rerun it with a single command.
3. **Model Versioning & Governance** – Register trained models, promote vetted versions to production, and keep a full audit trail.
4. **Deployment** – Serve models as REST APIs, batch jobs, or export them to cloud services (AWS SageMaker, Azure ML, GCP AI Platform) directly from the registry.
5. **End‑to‑End MLOps Integration** – Combine Tracking, Projects, and Registry with orchestration tools (Kubeflow, Airflow) to build CI/CD pipelines for ML.

### Integration with Common ML Workflows

| Workflow Step | How MLflow Fits In |
|---------------|---------------------|
| **Data preparation** | Log data‑version tags or dataset artifacts via Tracking. |
| **Model training** | Use `mlflow.start_run()` (or `mlflow.autolog()`) to capture parameters, metrics, and model artifacts automatically [2]. |
| **Packaging** | Wrap training code as an MLflow Project; define environment with `conda.yaml` or Dockerfile for repeatable runs [3]. |
| **Model packaging** | Save the trained object with `mlflow.<framework>.log_model()`, creating a portable “flavor” [4]. |
| **Model registry** | Register the model (`mlflow.register_model`) and move it through stages (Staging → Production) via the Model Registry UI or API [5]. |
| **Deployment** | Deploy the registered model to a REST endpoint, Spark cluster, or cloud‑managed service using built‑in deployment tools. |
| **Monitoring** | Track inference metrics by logging them back to MLflow Tracking, enabling drift detection and continuous evaluation. |

### Concise Summary

MLflow offers a **four‑module** framework—**Tracking**, **Projects**, **Models**, and **Model Registry**—that together enable reproducible experiments, standardized model packaging, robust version control, and seamless deployment. Organizations use it to accelerate MLOps pipelines, enforce governance, and collaborate on model development across diverse environments [1][2][3][4][5].

---
### Sources
[1] **MLflow Overview** – https://mlflow.org/docs/latest/index.html
[2] **MLflow Tracking Documentation** – https://mlflow.org/docs/latest/tracking.html
[3] **MLflow Projects Documentation** – https://mlflow.org/docs/latest/projects.html
[4] **MLflow Models Documentation** – https://mlflow.org/docs/latest/models.html
[5] **MLflow Model Registry Documentation** – https://mlflow.org/docs/latest/model-registry.html

Answering query using Neo4J vector DB

output/user_questions.md:

What have patients said about their quality of rest during their stay?

output/final_answer.md:

## Patient Feedback on Quality of Rest During Hospital Stay

Patients consistently report negative experiences with sleep quality during their hospital stay. The primary issues cited are:

1. **Physical discomfort** – uncomfortable beds that hinder restful sleep.
2. **Environmental noise** – constant machine beeping and nighttime interruptions.

### Key Review Excerpts
- **Crystal Johnson** (Smith, Edwards & Obrien): *"The uncomfortable beds made it difficult to get a good night’s sleep during my stay."*
- **Miranda Frey** (Brown‑Golden): *"The uncomfortable beds made it difficult to get a good night’s sleep during my stay."*
- **Jesse Tucker** (Wallace‑Hamilton): *"The constant interruptions during the night, however, affected my ability to rest."*
- **Heather Lewis** (Brown‑Golden): *"The constant beeping of machines in the ward was bothersome and made it difficult to get quality rest."*

### Overall Sentiment
Patients consistently report negative experiences with sleep quality during their hospital stay, citing physical discomfort and environmental noise as the main contributors.

### Sources
[1] HealthcareCypher: Patient Reviews – https://healthcarecypher.com/patient-reviews

output/user_questions.md:

What have patients said about hospital efficiency? Highlight details from specific reviews.

output/final_answer.md:

## Hospital Efficiency – Patient Review Highlights

Patients have commented on the efficiency of care at several hospitals. The following excerpts illustrate specific observations:

- **Brown Inc.** – *Jennifer McCall, Stephen Hernandez*: “The medical team was **efficient**.” This direct praise indicates that patients perceived the care process as swift and well‑coordinated.

- **Little‑Spencer** – *Shawn Ellis, Tracy Dalton*: “The well‑organized approach to my treatment contributed to a positive overall experience.” The phrase “well‑organized” implies streamlined processes and timely care, suggesting efficient workflow.

- **Smith, Edwards and Obrien** – *Nancy Nichols, Martin Gilbert*: “The hospital staff went above and beyond to make me comfortable during my stay. They even organized daily activities to keep patients engaged.” While this review focuses on comfort and engagement, it does not explicitly mention efficiency.

- **Smith, Edwards and Obrien** – *Tara Harris, Garrett Gomez*: “My experience at the hospital was positive overall. The medical team was competent, and the facilities were modern and well‑equipped.” This review highlights competence and modern facilities but lacks mention of efficiency.

These insights are drawn from patient reviews retrieved via the HealthcareReview tool.

### Sources

[1] HealthcareReview – Hospital efficiency patient reviews.

Answering question from Neo4J graph DB

output/user_questions.md:

Which physician has treated the most patients covered by Cigna?

output/final_answer.md:

The physician who has treated the most patients covered by Cigna is **Kayla Lawson**, who has seen **10 patients**.

### Sources
[1] HealthcareCypher: Physician Coverage Data – https://healthcarecypher.com/physician-coverage

output/user_questions.md:

What is the average visit duration for emergency visits in North Carolina?

output/final_answer.md:

## Average Emergency Visit Duration in North Carolina

The HealthcareCypher sub‑agent reports an average duration of **≈ 21,705 minutes** (about **15 days**) for emergency department visits in North Carolina. This figure likely reflects the total time patients spend in the hospital following an ED visit, rather than the time spent in the ED itself. No specific data source or time period was provided, so the number should be interpreted with caution.

### Sources

[1] HealthcareCypher: Average ED visit duration in North Carolina – 21,705 minutes (no source cited).

Accummulation

output/user_questions.md:

Which state had the largest percent increase in Medicaid visits from 2022 to 2023?

output/final_answer.md:

## Largest Percent Increase in Medicaid Visits (2022‑2023)

The HealthcareCypher sub‑agent reports that **Texas (TX)** had the largest percentage increase in Medicaid visits between 2022 and 2023, with an increase of **≈ 8.82 %**. This figure comes from the HealthcareCypher database’s Medicaid visit statistics by state for the two years.

### Sources

[1] HealthcareCypher: Medicaid visit statistics by state – 2022 and 2023.

output/user_questions.md:

Which payer provides the most coverage in terms of total billing amount?

output/final_answer.md:

## Payer with the Highest Total Coverage Amount

The payer that provides the most coverage in terms of total billing amount is **UnitedHealthcare**, with a total coverage amount of **$52,221,646.74**.

### Summary of Coverage Totals
| Payer | Total Coverage Amount |
|-------|------------------------|
| UnitedHealthcare | $52,221,646.74 |
| Cigna | $52,161,684.47 |
| Blue Cross | $51,440,502.96 |
| Medicaid | $50,464,204.50 |
| Aetna | $48,806,964.29 |

The figures above are derived from the healthcare database query on payer coverage totals.

### Source
[1] HealthcareCypher query result.

output/user_questions.md:

Categorize all patients' reviews into "Positive", "Neutral", and "Negative". Provide totals and percentage of the categories.

output/final_answer.md:

## Sentiment Categorization of Patient Reviews

The dataset contains four patient reviews, all of which are classified as **Positive**. No reviews were classified as **Neutral** or **Negative**.

| Sentiment | Count | Percentage |
|-----------|-------|------------|
| Positive | 4 | 100 % |
| Neutral | 0 | 0 % |
| Negative | 0 | 0 % |

### Source
[1] HealthcareCypher query result on patient reviews.

Answering question from Neo4J graph AND vector DB

output/user_questions.md:

Query the graph database to show me the reviews written by patient 7674

output/final_answer.md:

## Reviews Written by Patient 7674

The following table summarizes the review(s) authored by patient **7674** in the healthcare graph database:

| Review ID | Hospital | Physician | Review Text |
|-----------|----------|-----------|-------------|
| 644 | Jones, Brown and Murray | Sandra Porter | “The hospital provided exceptional care, but the billing process was confusing and frustrating. Clearer communication about costs would have been appreciated.” |

### Sources
[1] HealthcareCypher query “reviews written by patient 7674”

Email RAG Deep Agent

Sub-Agent:

Run:

Takes about 8 minutes to run.

$ uv run python -m src.EmailRAG.EmailRAG

Main Agent Outputs:

output/email_request_{timestamp}.md:

Escalation Criteria: There's an immediate risk of electrical, water, or fire damage
Escalation Dollar Criteria: 100000
Escalation Emails: brog@abc.com, bigceo@company.com

Date: Thu, 3 Apr 2026 11:36:10 +0000
From: City of Los Angeles Building and Safety Department <inspections@lacity.gov>
Reply-To: Admin <admin@building-safety.la.com>
To: West Coast Development <admin@west-coast-dev.com>
Cc: Donald Duck <donald@duck.com>, Support <inspections@lacity.gov>
Message-ID: <f967b2d6-1036-11f0-9701-9775a4ad682f@prod.outlook.com>
References: <7d29dafe-1037-11f0-a588-8f6b1b834703@prod.outlook.com> <859cc44e-1037-11f0-b15c-6b9d5cb20c47@prod.outlook.com>
Subject: Project 345678123 - Sunset Luxury Condominiums
Location: Los Angeles, CA

Following an inspection of your site at 456 Sunset Boulevard, we have identified the following building code violations:
Electrical Wiring: Exposed wiring was found in the underground parking garage, posing a safety hazard.
Fire Safety: Insufficient fire extinguishers were available across multiple floors of the structure under construction.
Structural Integrity: The temporary support beams in the eastern wing do not meet the load-bearing standards specified in local building codes.

Required Corrective Actions: Replace or properly secure exposed wiring to meet electrical safety standards.
Install additional fire extinguishers in compliance with fire code requirements. Reinforce or replace temporary support beams to ensure structural stability.
Deadline for Compliance: Violations must be addressed no later than December 31.
Failure to comply may result be a stop-work order and additional fines.
Contact: For questions or to schedule a re-inspection, please contact the Building and Safety Department at (555) 456-7890 or email inspections@lacity.gov.
---
## Key Findings

The email from **City of Los Angeles Building and Safety Department** concerning **Project 345678123 - Sunset Luxury Condominiums** at **456 Sunset Boulevard, Los Angeles, CA** outlines several critical building code violations that pose an immediate risk of electrical, water, or fire damage.

### Violations
[1] **Electrical Wiring**: Exposed wiring found in the underground parking garage, creating a safety hazard.
[2] **Fire Safety**: Insufficient fire extinguishers across multiple floors of the structure under construction.
[3] **Structural Integrity**: Temporary support beams in the eastern wing do not meet load‑bearing standards per local building codes.

### Required Corrective Actions
[1] Replace or properly secure exposed wiring to meet electrical safety standards.
[2] Install additional fire extinguishers in compliance with fire code requirements.
[3] Reinforce or replace temporary support beams to ensure structural stability.

### Deadline for Compliance
Violations must be addressed no later than **December 31, 2026**.

### Potential Penalties
Failure to comply may result in a stop‑work order and additional fines (amount not specified).

### Escalation Criteria
- Immediate risk of electrical, water, or fire damage.
- Escalation dollar threshold: **$100,000**.
- Escalation contacts: **brog@abc.com**, **bigceo@company.com**.

### Contact Information
Building and Safety Department: (555) 456‑7890, email: inspections@lacity.gov.

Sub-agent Outputs:

output/email_extract.md:

{
    "name": null,
    "phone": null,
    "email": null,
    "project_id": 345678123,
    "site_location": "456 Sunset Boulevard, Los Angeles, CA",
    "violation_types": [
        "Electrical Wiring",
        "Fire Safety",
        "Structural Integrity"
    ],
    "required_changes": [
        "Replace or properly secure exposed wiring to meet electrical safety standards.",
        "Install additional fire extinguishers in compliance with fire code requirements.",
        "Reinforce or replace temporary support beams to ensure structural stability."
    ],
    "max_potential_fine": null,
    "date_of_email": null,
    "compliance_deadline": null
}
This email warrants an escalation

LangSmith Application trace

https://smith.langchain.com/

Local log

/var/log/ragagent/log

Diagnostics

HTTP/3 curl:


$ docker run --rm ymuski/curl-http3 curl --http3 --verbose https://<nodeport service>:<nodeport>/health/ready

To build your own HTTP/3 curl: https://curl.se/docs/http3.html

Neo4J Graph DB

Sample Cypher Queries:

Visit with id:56:


MATCH (v:Visit) WHERE v.id = 56 RETURN v;

Which patient was involved in Visit id:56


MATCH (p:Patient)-[h:HAS]->(v:Visit) where v.id=56 return v,h,p

Which physician treated the patient i Visit id:56


MATCH (p:Patient)-[h:HAS]->(v:Visit)<-[t:TREATS]-(ph:Physician) where v.id=56 return v,h,p,t,ph

All relationships going in and out of Visit id:56


MATCH (v:Visit)-[r]-(n) where v.id=56 return v,r,n

Sample Cypher Accumulation:

Total visits and bill paid by payer Aetna in Texas:

MATCH (p:Payer)<-[c:COVERED_BY]-(v:Visit)-[:AT]->(h:Hospital)
WHERE p.name = "Aetna"
AND h.state_name = "TX"
RETURN COUNT(\*) as num_visits,
SUM(c.billing_amount) as total_billing_amount;

Coverage-Guided Fuzz Testing

To run fuzz-testing using Google's atheris and coverage:

$ uv run coverage run -m src.EmailRAG.fuzzer -atheris_runs=100

$ uv run coverage run -m src.rag_agent.fuzzer -atheris_runs=100

Generate HTML report and view it:

$ uv run python -m coverage html
$ (cd htmlcov && uv run python -m http.server 8000)

Name		Name	Last commit message	Last commit date
Latest commit History 407 Commits
.circleci		.circleci
data		data
images		images
migrations		migrations
src		src
test		test
.gitattributes		.gitattributes
.gitignore		.gitignore
.python-version		.python-version
Dockerfile		Dockerfile
Langchain.sql		Langchain.sql
LoadNeo4J.sh		LoadNeo4J.sh
README.md		README.md
alembic.ini		alembic.ini
build.sh		build.sh
env.py		env.py
hypercorn.sh		hypercorn.sh
hypercorn.toml		hypercorn.toml
langgraph.json		langgraph.json
pyproject.toml		pyproject.toml
pytest.ini		pytest.ini
uv.lock		uv.lock

Folders and files

Latest commit

History

Repository files navigation

LLM-RAG Deep Agent using LangChain, LangGraph, LangSmith

Infrastructure components:

Sources of information for RAG

Unstructured

Structured

Environment

Data Source Preparation:

PostgreSQL Database

Database Migration

PostgresSQL PGVector

Neo4J Graph DB

Local Model

Ollama

Google VertexAI / GenAI

Start the application:

HTTP/3 curl:

TLS certificates

Chrome browser

Home controller endpoints:

Hospital controller endpoints:

Launch LangGraph Server

Checkpointed RAG

RAG ReAct Agent

Run local Langgraph server

Main Agent:

Question & Answer RAG Deep Agent

Sub-Agent

RAG Deep Agent answers question from Vector and Graph Database

Answering questions from mocked-up API call:

Answering questions from Postgres Vector Store:

Answering query using Neo4J vector DB

Answering question from Neo4J graph DB

Accummulation

Answering question from Neo4J graph AND vector DB

Email RAG Deep Agent

Sub-Agent:

Run:

Main Agent Outputs:

Sub-agent Outputs:

LangSmith Application trace

Local log

Diagnostics

Neo4J Graph DB

Sample Cypher Queries:

Sample Cypher Accumulation:

Coverage-Guided Fuzz Testing

About

Topics

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Uh oh!

Contributors

Uh oh!

Languages

Packages