search_guides · Zero-cost RAG

How keyword RAG works

The agent's search_guides tool fetches a remote design manual, splits it into paragraphs, ranks them by keyword overlap, and returns the top 3 matches — no embeddings, no vector DB.

1. Fetch

Download raw markdown

2. Chunk

Split by paragraph

3. Tokenize

Clean query words

4. Score

Keyword overlap

5. Retrieve

Top 3 sections

Fetch — ingest remote markdown

The pipeline starts with a plain HTTP fetch of the raw GitHub file. No preprocessing — just the full markdown text in memory.

GET https://raw.githubusercontent.com/RayFernando1337/llm-cursor-rules/main/fire-your-design-team.md

Bytes fetched

10,420

Lines

282

Source

GitHub Raw

fire-your-design-team.md

Chunk — split by double newline

Markdown is split on \n\n+ (paragraph boundaries). Short fragments like --- are dropped, but markdown headings (e.g. ## Typography System) are kept — they label sections and help topic matching.

Raw splits

Kept

paragraphs + headings

Dropped

short noise only

Kept chunks (sample)

chunk #053 chars

# shadcn/ui with Tailwind v4 Design System Guidelines

chunk #1231 chars

This document outlines design principles and implementation guidelines for applications using shadcn/ui with Tailwind v4. These guidelines ensure consistency, a…

chunk #225 chars

## Core Design Principles

chunk #3385 chars

### 1. Typography System: 4 Sizes, 2 Weights - **4 Font Sizes Only**: - Size 1: Large headings - Size 2: Subheadings/Important content - Size 3: Body text…

chunk #4275 chars

### 2. 8pt Grid System - **All spacing values must be divisible by 8 or 4** - **Examples**: - Instead of 25px padding → Use 24px (divisible by 8) - Instead …

chunk #5244 chars

### 3. 60/30/10 Color Rule - **60%**: Neutral color (white/light gray) - **30%**: Complementary color (dark gray/black) - **10%**: Main brand/accent color (e.g.…

Dropped as noise (if any)

All splits in this document were kept — section headings now count as valid chunks.

Visual: each blank line pair in the source becomes a split point. Short fragments between splits never reach scoring.

Tokenize — normalize query words

The query string is lowercased and split on non-alphanumeric characters. Duplicates in the query count once when scoring.

Example query

"typography font sizes weights"

Tokens

typographyfontsizesweights

Auto-Layout → auto, layout · flex-wrap: nowrap → flex, wrap, nowrap

Score — keyword intersection

Each chunk gets a score equal to how many unique query tokens appear in that chunk. More overlapping keywords = higher rank.

6 chunks matchedMin score: 1

typographyfontsizesweights

#1 in context

### 1. Typography System: 4 Sizes, 2 Weights - **4 Font Sizes Only**: - Size 1: Large headings - Size 2: Subheadings/Important content - Siz…

typographyfontsizesweights

#2 in context

### Core Design Principles - [ ] Typography: Uses only 4 font sizes and 2 font weights (Semibold, Regular) - [ ] Spacing: All spacing values…

fontsizesweights

#3 in context

### Font Sizes & Weights - **Strictly limit to 4 distinct sizes**: - Size 1: Large headings (largest) - Size 2: Subheadings - Size 3: Body t…

fontsizes

### Common Issues to Flag - [ ] Too many font sizes (more than 4) - [ ] Inconsistent spacing values (not divisible by 8 or 4) - [ ] Overuse …

typography

## Typography System

typography

### Typography Implementation - **Reference shadcn's typography primitives** for consistent text styling - **Use monospace variant** for num…

Retrieve — top 3 joined context

Matching chunks are sorted by score, filtered by the minimum threshold, limited to 3, then joined with \n\n---\n\n. This string is injected into the agent prompt as grounded context.

### 1. Typography System: 4 Sizes, 2 Weights
- **4 Font Sizes Only**:
  - Size 1: Large headings
  - Size 2: Subheadings/Important content
  - Size 3: Body text
  - Size 4: Small text/labels
- **2 Font Weights Only**:
  - Semibold: For headings and emphasis
  - Regular: For body text and general content
- **Consistent Hierarchy**: Maintain clear visual hierarchy with limited options

---

### Core Design Principles
- [ ] Typography: Uses only 4 font sizes and 2 font weights (Semibold, Regular)
- [ ] Spacing: All spacing values are divisible by 8 or 4
- [ ] Colors: Follows 60/30/10 color distribution (60% neutral, 30% complementary, 10% accent)
- [ ] Structure: Elements are logically grouped with consistent spacing

---

### Font Sizes & Weights
- **Strictly limit to 4 distinct sizes**:
  - Size 1: Large headings (largest)
  - Size 2: Subheadings
  - Size 3: Body text
  - Size 4: Small text/labels (smallest)
- **Only use 2 font weights**:
  - Semibold: For headings and emphasis
  - Regular: For body text and most UI elements
- **Common mistakes to avoid**:
  - Using more than 4 font sizes
  - Introducing additional font weights
  - Inconsistent size application

Short query (≤ 4 tokens)

minScore = 1 — any single keyword match qualifies.

Long query (≥ 5 tokens)

minScore = 2— prevents one common word (e.g. "primary") from returning unrelated sections.

Why this approach?

Zero costNo embedding API calls or vector database — just string matching.
DeterministicSame query always returns the same chunks. Easy to debug and demo.
GroundedAgent recommendations cite real paragraphs from the design manual.