Examples

Deep Research

Use Constclaw's Expand & Elaborate mode to dive deep into complex topics. This example shows how a short passage can be expanded into a comprehensive research briefing.

Source

🌐 arxiv.org 📖 Research paper abstract

Selected text

"Retrieval-Augmented Generation (RAG) combines information retrieval with language model generation. The retriever fetches relevant documents from an external knowledge base, and the generator conditions its output on both the query and retrieved context. RAG has shown improvements over purely parametric models on knowledge-intensive NLP tasks."

📖 Expand & Elaborate

OpenClaw Response

Retrieval-Augmented Generation (RAG) — Deep Dive

Background

RAG was introduced by Lewis et al. (2020) at Facebook AI Research. The core insight: large language models store knowledge in their parameters, but this knowledge is static, limited by training data, and expensive to update. RAG solves this by adding a retrieval step that fetches current, relevant documents at inference time.

Architecture

A RAG system has two main components:

Retriever — Typically a dense passage retriever (DPR) that encodes both the query and documents into vector embeddings, then finds the top-k most similar documents using approximate nearest neighbor (ANN) search. Common tools: FAISS, Pinecone, Weaviate, ChromaDB.
Generator — A seq2seq model (originally BART, now commonly GPT-4, Claude, etc.) that receives the concatenation of the query + retrieved passages as input and generates the final answer.

RAG vs. fine-tuning

Aspect

RAG

Fine-tuning

Knowledge updates

Instant (update docs)

Requires retraining

Hallucination risk

Lower (grounded)

Higher

Cost

Retrieval overhead

Training compute

Traceability

Can cite sources

No source attribution

Advanced techniques

Hybrid search — Combine dense (semantic) and sparse (BM25 keyword) retrieval for better recall
Re-ranking — Use a cross-encoder to re-score retrieved passages before generation
Chunking strategies — Split documents by semantic boundaries rather than fixed token windows
Query decomposition — Break complex queries into sub-queries, retrieve for each, then synthesize
Self-RAG — The model decides when to retrieve and critically evaluates the retrieved context

Real-world applications

RAG powers customer support chatbots (retrieval over help docs), enterprise search (internal knowledge bases), legal research (case law retrieval), and medical Q&A (retrieval over PubMed). It's the dominant architecture for production LLM applications that need factual accuracy and up-to-date knowledge.

Key papers

Lewis et al. (2020) — "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks"
Karpukhin et al. (2020) — "Dense Passage Retrieval for Open-Domain Question Answering"
Asai et al. (2023) — "Self-RAG: Learning to Retrieve, Generate, and Critique"

Tip: The Expand mode works best with short, dense passages — definitions, abstracts, or topic sentences. The AI uses its training knowledge to elaborate far beyond the source text.

Try it yourself

Select any paragraph from a research paper or technical article, choose Expand & Elaborate, and click Analyze. Constclaw will generate a comprehensive briefing with background, comparisons, and further reading.

← All Examples Translation →