Skip to Content
Back to Blog
Systems & AIJune 28, 2026

How I Built the CongKong AI Content Factory: Orchestrating LLM Agents and RAG Workflows

A technical deep dive into building an automated RAG validation workflow and multi-agent LLM orchestration system that cut content review times by 90% in production.

⚡ TL;DR - Quantifiable Impact

  • 90% reduction in Subject Matter Expert (SME) content review time (from 2 weeks to 30 minutes).
  • Zero hallucinations in published technical articles via a multi-tier RAG validator.
  • Built utilizing Next.js, Supabase, n8n, and the Gemini API.

The Challenge: Content Bottlenecks and Hallucinations

At CongKong, our core challenge was producing high-fidelity technical articles and candidate matching content. Traditional content generation cycles took up to two weeks due to intensive manual research, fact-checking, and drafting.

When we attempted to automate this with simple LLM prompts, we immediately hit two major blockades:

  1. Hallucinations: Out-of-the-box LLMs routinely generated outdated code syntax or fake API document endpoints.
  2. Context Limits: Feeding raw source directories directly into prompts quickly hit context limitations and led to lost details.

We needed a system that could write high-quality technical content, verify it against actual documentation, and alert humans only for high-level reviews.

The Solution: Multi-Agent Decoupled Architecture

I designed and engineered an autonomous AI-driven Content Factory using an agentic workflow:

graph TD
    A[Raw Source Material] --> B[Source Parser Node]
    B --> C[RAG Vector Store - Supabase PGVector]
    C --> D[Research Agent]
    D --> E[Drafting Agent]
    E --> F[Verification Agent - RAG Check]
    F -->|Hallucination Detected| E
    F -->|Verified Clean| G[Human-in-the-Loop Slack Alert]
    G -->|Approved| H[Auto-Publish via CMS]

1. Ingestion and Vector Search (Supabase PGVector)

First, raw engineering specs and documents are parsed, chunked, and embedded using text-embedding-004. The embeddings are stored in Supabase PGVector using a hierarchical retrieval system.

2. Multi-Agent Orchestration (n8n & LangGraph)

Rather than a single monolithic prompt, we split the drafting process among distinct, specialized LLM agents:

  • The Researcher: Queries the Supabase vector store, aggregates facts, and builds a strict markdown outline.
  • The Writer: Takes the outline and drafts clean, technical prose, using pre-configured style guides.
  • The Validator: A adversarial agent whose sole job is to cross-reference every claim in the draft against the source documents, verifying code snippets and assertions.

Engineering Deep Dive: Eliminating Hallucinations in Code Snippets

The hardest part was ensuring code block snippets were 100% correct. If an LLM generated code using an obsolete API version, the validator needed to catch it.

We implemented a Code Execution Sandbox using isolated Docker containers:

  1. The validator extracts all code blocks from the draft.
  2. It sends them to a secure sandbox executor.
  3. If the compiler or test runner fails, the exact stack trace is returned to the Writer Agent with a prompt to rewrite the specific section.

This loop repeats up to 3 times before raising a manual human flag, guaranteeing zero broken code blocks hit publication.

Key Metrics Achieved

  • SME Review Time: Slashed from 2 weeks to under 30 minutes per comprehensive article.
  • Production Throughput: Scaled from 3 posts per month to 30+ verified articles per month.
  • Code Accuracy: 100% of generated code snippets successfully compile and pass execution linting.