> ## Documentation Index
> Fetch the complete documentation index at: https://docs.insitechat.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Concepts — How InsiteChat Works Under the Hood

> InsiteChat technical foundation: RAG, embeddings, hybrid search with Reciprocal Rank Fusion, system prompts, and vector databases. Plain-English explainers.

**InsiteChat** is built on retrieval-augmented generation (RAG) with a hybrid retrieval layer. The pages below explain — in plain English — how each piece works and why we made the architectural choices we did. Read them in order if you're new to RAG; skim individually if you're evaluating specific aspects.

## The five core concepts

<CardGroup cols={2}>
  <Card title="What is RAG?" icon="brain-circuit" href="/concepts/what-is-rag">
    Retrieval-Augmented Generation — how InsiteChat grounds AI answers in your content instead of relying on the LLM's frozen pre-training. The technique that makes chatbots factually accurate.
  </Card>

  <Card title="Embeddings" icon="vector-square" href="/concepts/embeddings">
    How text becomes high-dimensional vectors that capture meaning. The foundation of semantic search — finds "cancel my subscription" matches "close my account" even though they share zero keywords.
  </Card>

  <Card title="Hybrid Search" icon="layer-group" href="/concepts/hybrid-search">
    Why InsiteChat combines vector (semantic) and BM25 (keyword) search with Reciprocal Rank Fusion. Catches edge cases — error codes, currency symbols, brand names — that vector-only systems miss.
  </Card>

  <Card title="System Prompts" icon="message-pen" href="/concepts/system-prompts">
    The standing instruction that defines your chatbot's persona, tone, and refusal behavior. One small block of text controls every response. InsiteChat ships 6 templates plus persona shaping.
  </Card>

  <Card title="Vector Databases" icon="database" href="/concepts/vector-databases">
    Where embeddings live. InsiteChat runs on pgvector — PostgreSQL with the vector extension — for single-source-of-truth storage, ACID guarantees, and sub-15ms nearest-neighbor search at scale.
  </Card>
</CardGroup>

## How the pieces fit together

A single query flowing through InsiteChat touches all five concepts:

1. Your content was previously chunked, **embedded**, and stored in the **vector database** at training time.
2. A visitor types a question. Their question is also **embedded**.
3. **Hybrid search** runs vector + BM25 retrieval and merges the rankings with RRF, returning the top-K most relevant chunks.
4. The retrieved chunks plus your **system prompt** plus conversation history are sent to the LLM.
5. The LLM generates an answer grounded in the retrieved context — this is **RAG**.

The whole loop runs in 1-3 seconds and produces an answer that cites the source pages it came from.

## Why these architectural choices

* **RAG over fine-tuning**: lets you update knowledge by editing content (minutes, free) instead of retraining a model (hours, expensive). See [What is RAG?](/concepts/what-is-rag) § "RAG vs fine-tuning".
* **Hybrid search over vector-only**: catches edge cases (error codes, GST numbers, ₹ symbols, technical terms) that pure semantic search misses. See [Hybrid Search](/concepts/hybrid-search).
* **pgvector over a separate vector DB**: single source of truth with chatbot config, conversations, and leads. No two-store sync problems. See [Vector Databases](/concepts/vector-databases).
* **Custom Q\&A pairs override retrieval**: precise control on high-stakes answers (pricing, refunds, hours) without losing the flexibility of RAG. See [Custom Q\&A](/training/custom-qa).

## Related operational topics

* [Website Crawler](/training/website-crawler) — how InsiteChat ingests content
* [Document Upload](/training/document-upload) — file formats, OCR, chunking
* [Syncing Content](/training/syncing-content) — keeping the chatbot current
* [Custom Q\&A](/training/custom-qa) — overriding AI-generated answers