gorag

A Go library for building Retrieval-Augmented Generation (RAG) flows with:

• knowledge loading and chunking
• embedding and response caching
• Qdrant vector store bootstrap helpers
• dense retrieval with optional reranking

Module path: github.com/raykavin/gorag

Packages

• engine: RAG pipeline building blocks
• cache: in-memory embedding and response caches
• pkg/store: Qdrant bootstrap and vector store setup helpers
• pkg/models: shared data models (KnowledgeChunk, ChatResponse, ToolCallLog)

Install

go get github.com/raykavin/gorag

Quick Start

package main

import (
	"context"
	"log"
	"time"

	"github.com/raykavin/gorag/cache"
	"github.com/raykavin/gorag/engine"
	"github.com/raykavin/gorag/pkg/store"
	"github.com/tmc/langchaingo/embeddings"
)

func main() {
	ctx := context.Background()

	// Create your base embedder (provider-specific).
	var base embeddings.Embedder

	// Optional embedding cache wrapper.
	embCache := cache.NewEmbedding(30*time.Minute, 10*time.Minute)
	cachedEmbedder, err := engine.NewCacheEmbedder(base, embCache)
	if err != nil {
		log.Fatal(err)
	}

	// Probe vector size from your chunks.
	vectorSize, err := store.ProbeEmbeddingSize(ctx, cachedEmbedder, nil)
	if err != nil {
		log.Fatal(err)
	}

	// Ensure Qdrant collection exists.
	_, err = store.PrepareCollection(
		ctx,
		"http://localhost:6333",
		"",          // api key
		"knowledge", // collection
		vectorSize,
		false,       // reIndexOnBoot
	)
	if err != nil {
		log.Fatal(err)
	}

	// Build vector store.
	vs, err := store.NewQdrant("http://localhost:6333", "", "knowledge", cachedEmbedder)
	if err != nil {
		log.Fatal(err)
	}

	// Build RAG engine.
	rag := engine.NewRAGEngine(vs, nil, embCache, engine.Config{
		TopK:          3,
		MinScore:      0.7,
		ForceAlways:   true,
		RetrievalMode: "dense", // or "reranker"
		RerankerK:     8,
	})

	contextBlock, used, err := rag.Query(ctx, "What is our refund policy?")
	if err != nil {
		log.Fatal(err)
	}
	if used {
		log.Println(contextBlock)
	}
}

Knowledge Loading

Use engine.ReadKnowledge(path) to read and expand JSON knowledge files.

Supported sources:

• inline chunk content
• explicit file list in files
• directory traversal via files_path

Supported file extensions:

• .txt, .md, .markdown, .csv

Example input shape (pkg/models.KnowledgeChunk):

[
  {
    "id": "KB_001",
    "category": "refunds",
    "files": ["docs/refunds.md"], // Load files from list
  },
  {
	"id": "KB_002",
    "category": "terms_and_conditions",
    "content": "The terms and conditions are...", // Inline chunk
  },
  {
    "id": "KB_003",
	"category": "faq",
    "files_path": "docs/faq" // Load all files in directory
  }
]

Retrieval Modes

engine.RAGEngine supports:

• dense:

  • similarity search with score threshold (MinScore)
  • optional fallback search without threshold if ForceAlways=true

• reranker:

  • dense candidate retrieval (RerankerK)
  • optional rerank step to return top TopK

Reranker

Use engine.NewCrossEncoderReranker(engine.CrossEncoderConfig{...}).

Important options:

• BaseURL: reranker API base URL (required)
• EndpointPath: defaults to /api/rerank
• Model: reranker model name (required)
• APIKey: optional (sent as Authorization: Bearer ... and api-key)
• Timeout: defaults to 20s

Cache

Embedding Cache

cache.Embedding:

• key: SHA-256(text)
• methods: Get, Set, Stats

Response Cache

cache.ResponseCache:

• key: normalized (text + historyKey) hash
• methods: Get, Set, Invalidate, Stats
• bounded size with oldest-item eviction
• TTL policy based on tool names used in ChatResponse.ToolCalls

Qdrant Helpers

pkg/store provides:

• PrepareCollection: create/check/delete collection on boot
• NewQdrant: build langchaingo Qdrant vector store
• ProbeEmbeddingSize: infer vector dimension from embedder output

Error Handling

Packages expose typed errors for common validation/setup failures (for example, nil embedders, empty URLs, invalid reranker config). Prefer errors.Is checks when branching.

Testing

Run all tests:

go test ./...

If your environment restricts default Go cache paths, run:

GOCACHE=/tmp/go-build go test ./...

---

Contributing

Contributions to gorag are welcome! Here are some ways you can help improve the project:

• Report bugs and suggest features by opening issues on GitHub
• Submit pull requests with bug fixes or new features
• Improve documentation to help other users and developers
• Share your custom strategies with the community

License

gorag is distributed under the MIT License.
For complete license terms and conditions, see the LICENSE file in the repository.

---

Contact

For support, collaboration, or questions about gorag:

Email: raykavin.meireles@gmail.com
GitHub: @raykavin