In the previous article, we discussed how to build a RAG system pipeline in Laravel.

Now that we already store embeddings in PostgreSQL, the next step is implementing semantic search.

Semantic search allows the AI system to understand meaning instead of exact keywords.

Traditional search:

textCopy
"authentication"

Semantic search:

textCopy
"how users login securely"

Both can return similar results.

Understanding Vector Similarity

Vector similarity measures how close two embeddings are.

Common metrics:

Most RAG systems use cosine similarity.

Creating Semantic Search Query

Generate query embedding:

phpCopy
<?php

$queryEmbedding = app(EmbeddingService::class)
    ->embed($query);

Search similar documents:

phpCopy
<?php

$embeddingString = '[' . implode(',', $queryEmbedding) . ']';

$documents = DB::select(
    "
    SELECT
        id,
        title,
        content,
        embedding <=> ?::vector AS distance
    FROM documents
    ORDER BY distance ASC
    LIMIT 5
    ",
    [$embeddingString]
);

What is <=> Operator?

pgvector provides special operators.

Lower distance means more similar.

Creating Search API Endpoint

Routes:

phpCopy
<?php

use App\Http\Controllers\SearchController;

Route::post('/search', SearchController::class);

Controller:

phpCopy
<?php

namespace App\Http\Controllers;

use App\Services\EmbeddingService;
use Illuminate\Http\Request;
use Illuminate\Support\Facades\DB;

class SearchController extends Controller
{
    public function __invoke(Request $request)
    {
        $query = $request->query;

        $embedding = app(EmbeddingService::class)
            ->embed($query);

        $embeddingString = '[' . implode(',', $embedding) . ']';

        $documents = DB::select(
            "
            SELECT
                id,
                title,
                content,
                embedding <=> ?::vector AS distance
            FROM documents
            ORDER BY distance ASC
            LIMIT 5
            ",
            [$embeddingString]
        );

        return response()->json($documents);
    }
}

Optimizing Vector Search Performance

Without indexes, vector search becomes slow.

Create vector index:

sqlCopy
CREATE INDEX documents_embedding_idx
ON documents
USING ivfflat (embedding vector_cosine_ops)
WITH (lists = 100);

Analyze table:

sqlCopy
ANALYZE documents;

Understanding Chunking

Large documents should be split into smaller chunks.

Bad approach:

textCopy
1 huge PDF → 1 embedding

Good approach:

textCopy
1 PDF → multiple chunks → multiple embeddings

Benefits:

• Better retrieval
• More accurate context
• Faster search
• Reduced hallucination

Example Chunking Strategy

phpCopy
<?php

$text = file_get_contents($path);

$chunks = str($text)
    ->split(1000);

Store each chunk separately.

Building a Context Builder

After retrieving documents, combine them into context.

phpCopy
<?php

$context = collect($documents)
    ->pluck('content')
    ->join("\n\n");

Example:

textCopy
You are an AI assistant.

Use the following context:

[DOCUMENTS]

Question:
How does Laravel queue work?

Preventing Hallucinations

One major advantage of RAG is reducing hallucinations.

Good prompt:

textCopy
Only answer using the provided context.
If the answer does not exist, say you do not know.

This dramatically improves reliability.

Building a Retrieval Service

phpCopy
<?php

namespace App\Services;

use Illuminate\Support\Facades\DB;

class RetrievalService
{
    public function search(string $query): array
    {
        $embedding = app(EmbeddingService::class)
            ->embed($query);

        $embeddingString = '[' . implode(',', $embedding) . ']';

        return DB::select(
            "
            SELECT
                title,
                content,
                embedding <=> ?::vector AS distance
            FROM documents
            ORDER BY distance ASC
            LIMIT 5
            ",
            [$embeddingString]
        );
    }
}

Next Part

In the next article, we will build:

• Full AI chat endpoint
• Streaming responses
• Chat memory
• Conversation history
• Production optimization
• Queue workers
• Background embedding jobs
• Multi-model architecture