# Local Vector Store (RAG Development)
Test RAG (Retrieval-Augmented Generation) locally without setting up an external
vector database. Documents are embedded and stored in-memory for fast iteration.
## How Local RAG Works
```
┌─────────────────────────────────────────────────────────────────┐
│ LOCAL RAG PIPELINE │
│ │
│ STEP 1: INDEX (one-time) STEP 2: RETRIEVE (per query) │
│ ───────────────────────── ──────────────────────────── │
│ │
│ "The Matrix is a 1999..." "What's a good sci-fi film?" │
│ │ │ │
│ ▼ ▼ │
│ ┌──────────────┐ ┌──────────────┐ │
│ │ Embedder │ │ Embedder │ │
│ │ (Vertex AI) │ │ (Vertex AI) │ │
│ └──────┬───────┘ └──────┬───────┘ │
│ │ │ │
│ ▼ ▼ │
│ ┌──────────────┐ ┌──────────────┐ │
│ │ Local Vector │◄─── compare ────│ Query Vector │ │
│ │ Store │ └──────────────┘ │
│ │ (in-memory) │────────────────────────────────► │
│ └──────────────┘ "The Matrix" (0.95 match) │
│ "Inception" (0.89 match) │
└─────────────────────────────────────────────────────────────────┘
```
## Features Demonstrated
| Feature | Flow / API | Description |
|---------|-----------|-------------|
| Local Vector Store | `define_dev_local_vector_store()` | In-memory store, no external DB needed |
| Document Indexing | `index_documents` | Embed and store film descriptions |
| Similarity Retrieval | `retrieve_documents` | Find semantically similar documents |
| Document Creation | `Document.from_text()` | Convert plain text to Genkit documents |
| Vertex AI Embeddings | `gemini-embedding-001` | Cloud-based embedding model |
## ELI5: Key Concepts
| Concept | ELI5 |
|---------|------|
| **RAG** | AI looks up your docs before answering — fewer hallucinations! |
| **Vector Store** | A database that finds "similar" items by meaning — "happy" finds docs about "joyful" too |
| **Local Store** | Runs on your computer, no cloud DB needed — perfect for testing before production |
| **Indexing** | Adding documents to the store — like a librarian cataloging new books |
| **Retrieval** | Finding documents that match a query — "sci-fi films" returns The Matrix, Inception |
## Quick Start
```bash
export GCLOUD_PROJECT=your-project-id
gcloud auth application-default login
./run.sh
```
Then open the Dev UI at http://localhost:4000.
## Setup
### 1. Authenticate with Google Cloud
Vertex AI embeddings require GCP authentication:
```bash
gcloud auth application-default login
```
### 2. Set Your Project ID
```bash
export GCLOUD_PROJECT=your-project-id
```
Or the demo will prompt you interactively.
### 3. Run the Sample
```bash
./run.sh
```
Or manually:
```bash
genkit start -- uv run src/main.py
```
## Testing This Demo
1. **Open DevUI** at http://localhost:4000
2. **Index documents first**:
- [ ] Run `index_documents` — indexes 10 classic film descriptions
3. **Test retrieval**:
- [ ] Run `retrieve_documents` — queries for "sci-fi film"
- [ ] Verify The Matrix, Inception, Star Wars appear as top matches
4. **Expected behavior**:
- Documents are embedded using Vertex AI (`gemini-embedding-001`)
- Vector store is in-memory (no external database)
- Retrieval returns semantically similar documents (top 3)
- No external vector database required
## Sample Documents
The sample indexes 10 classic films including The Godfather, The Dark Knight,
Pulp Fiction, Inception, The Matrix, Star Wars, and more.
## Development
The `run.sh` script uses `watchmedo` for hot reloading on file changes.
