Entry point: /faion-net — invoke this skill for automatic routing to the appropriate domain.

RAG Engineer Skill

Communication: User's language. Code: English.

Purpose

Specializes in RAG (Retrieval Augmented Generation) systems. Covers document processing, embeddings, vector search, and retrieval optimization.

Context Discovery

Auto-Investigation

Check these project signals before asking questions:

Signal	Where to Check	What to Look For
Dependencies	package.json, requirements.txt	langchain, llamaindex, qdrant-client, chromadb, weaviate-client
Vector DB	docker-compose.yml, .env	Qdrant, Weaviate, Chroma config/containers
Document dirs	/docs, /data, /content	Documents to index (PDF, MD, TXT)
Existing embeddings	Grep for "embed", "vector", "retriever"	Current RAG implementation

Discovery Questions

question: "What's your RAG use case?"
header: "RAG Goal"
multiSelect: false
options:
  - label: "Documentation Q&A"
    description: "Answer questions from internal docs"
  - label: "Knowledge base search"
    description: "Semantic search over articles/guides"
  - label: "Code search/retrieval"
    description: "Find relevant code snippets"
  - label: "Customer support"
    description: "Context-aware support responses"

question: "Which vector database?"
header: "Vector DB"
multiSelect: false
options:
  - label: "Qdrant (recommended for production)"
    description: "Fast, scalable, rich filtering"
  - label: "Chroma (recommended for dev/prototyping)"
    description: "Simple, local, easy setup"
  - label: "Weaviate (for knowledge graphs)"
    description: "Hybrid search, graph features"
  - label: "pgvector (for PostgreSQL projects)"
    description: "Vector extension for existing Postgres"

question: "Document volume and type?"
header: "Data Characteristics"
multiSelect: false
options:
  - label: "Small (<1000 docs, mostly text)"
    description: "Simple chunking sufficient"
  - label: "Medium (1000-10000 docs)"
    description: "Consider hybrid search + reranking"
  - label: "Large (>10000 docs, mixed formats)"
    description: "Advanced chunking + metadata filtering"
  - label: "Code repository"
    description: "AST-aware chunking needed"

question: "Do you need hybrid search (vector + keyword)?"
header: "Search Strategy"
multiSelect: false
options:
  - label: "Yes - combine semantic + exact matching"
    description: "Hybrid search for best results"
  - label: "No - semantic search only"
    description: "Vector similarity sufficient"

Scope

Area	Coverage
Chunking	Text splitting, semantic chunking, overlap strategies
Embeddings	Text vectorization, similarity search, models
Vector DBs	Qdrant, Weaviate, Chroma, pgvector
Retrieval	Hybrid search, reranking, metadata filtering
RAG Systems	Architecture, evaluation, agentic RAG

Quick Start

Task	Files
Basic RAG	chunking-basics.md → embedding-basics.md → rag-architecture.md
Vector DB setup	db-comparison.md → db-qdrant.md (recommended)
Advanced retrieval	hybrid-search-basics.md → reranking-basics.md
RAG evaluation	rag-eval-metrics.md → rag-eval-methods.md
Agentic RAG	agentic-rag.md

Methodologies (22)

Chunking (2):

• chunking-basics: Size, overlap, delimiters
• chunking-advanced: Semantic, recursive, custom

Embeddings (4):

• embedding-basics: Fundamentals, similarity
• embedding-generation: API usage, batching
• embedding-models: Comparison, selection
• embedding-applications: Use cases, patterns

Vector Databases (4):

• db-comparison: Feature comparison, selection
• db-qdrant: Setup, indexing, search (recommended)
• db-weaviate: Knowledge graphs, hybrid search
• db-chroma: Local dev, prototyping
• vector-database-setup: General setup patterns

Retrieval (4):

• hybrid-search-basics: Vector + keyword search
• hybrid-search-implementation: Production patterns
• reranking-basics: Cross-encoder fundamentals
• reranking-models: Cohere, MixedBread, custom

RAG Systems (7):

• rag: RAG overview, fundamentals
• rag-architecture: System design, components
• rag-implementation: Production patterns
• rag-eval-metrics: Relevance, faithfulness, correctness
• rag-eval-methods: Evaluation frameworks
• agentic-rag: Agent-driven retrieval
• graph-rag-advanced-retrieval: Knowledge graphs

Architecture

Document Ingestion
    ↓
Chunking (semantic/fixed)
    ↓
Embedding Generation
    ↓
Vector Database Storage
    ↓
Query Processing
    ↓
Retrieval (vector + hybrid)
    ↓
Reranking
    ↓
Context Assembly
    ↓
LLM Generation

Code Examples

Basic RAG Pipeline

from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_openai import OpenAIEmbeddings
from langchain_chroma import Chroma

# Chunk documents
splitter = RecursiveCharacterTextSplitter(
    chunk_size=1000,
    chunk_overlap=200
)
chunks = splitter.split_documents(docs)

# Generate embeddings and store
embeddings = OpenAIEmbeddings()
vectorstore = Chroma.from_documents(chunks, embeddings)

# Retrieve
retriever = vectorstore.as_retriever(
    search_type="similarity",
    search_kwargs={"k": 5}
)
results = retriever.invoke("query")

Hybrid Search with Qdrant

from qdrant_client import QdrantClient
from qdrant_client.models import Distance, VectorParams, Filter

client = QdrantClient("localhost", port=6333)

# Create collection
client.create_collection(
    collection_name="docs",
    vectors_config=VectorParams(size=1536, distance=Distance.COSINE)
)

# Hybrid search
results = client.search(
    collection_name="docs",
    query_vector=query_embedding,
    query_filter=Filter(...),
    limit=10
)

Reranking

from cohere import Client

co = Client(api_key="...")

# Rerank retrieved docs
reranked = co.rerank(
    query="query text",
    documents=[doc.text for doc in results],
    top_n=3,
    model="rerank-english-v3.0"
)

Evaluation Metrics

Metric	Measures
Retrieval Precision	Relevant docs in results
Retrieval Recall	Coverage of relevant docs
MRR	Mean reciprocal rank
NDCG	Ranking quality
Faithfulness	Grounding in context
Answer Relevance	Response matches query

Related Skills

Skill	Relationship
faion-llm-integration	Uses embedding APIs
faion-ai-agents	Agentic RAG patterns
faion-ml-ops	RAG evaluation

---

RAG Engineer v1.0 | 22 methodologies