Recallr

README

Recallr Chat Context Example

This example demonstrates how to run an end-to-end chat interaction using the Recallr Golang library with:

• ✅ DragonflyDB as a Redis-compatible memory store
• 🧠 Long-term user memory
• 💬 Session context tracking
• ✍️ Customizable prompt injection
• 🧩 Pluggable LLM clients (example uses OpenAI-compatible client)

---

🔧 Prerequisites

1. Install Docker
2. Run DragonflyDB locally

docker run -d --name dragonfly -p 6379:6379 docker.dragonflydb.io/dragonflydb/dragonfly

1. Clone Recallr

git clone https://github.com/Jsee98/Recallr
cd Recallr

1. Run Example

go run examples/main.go

---

📦 What It Does

• Starts a Dragonfly Redis instance.
• Initializes session manager, user memory, and prompt manager.
• Stores user facts into long-term memory.
• Builds a prompt using system prompt + user memory + chat history + current input.
• Sends the prompt to an LLM client (here: OpenAI-compatible streaming client).
• Streams the assistant's response line-by-line.
• Stores both user input and assistant output in session memory.

---

🧠 Message Roles Explained

Role	Description
system	Sets the LLM's behavior, tone, and persona
user	Contains either a starting prompt or actual user input
assistant	LLM response, either full or streamed

---

🧙‍♂️ Prompt System

Prompt configuration is done via environment variables:

• RECALLR_SYSTEM_PROMPT: Instructions for the LLM’s behavior (required)
• RECALLR_USER_PROMPT: Optional user-level bootstrap message

These are injected during prompt building and can be reloaded at runtime:

promptMgr.Reload()

---

🧠 Fact Summarizer

Recallr supports optional summarization of long-term user memory.

If your LLM client implements:

type FactSummarizer interface {
    Summarizer(ctx context.Context, facts map[string]string) (string, error)
}

Then this summarizer is used to compress user facts before injecting into the prompt.

Otherwise, the default summarizer generates a simple readable block:

User Facts:
- location: Berlin
- language: Go

---

🔌 LLM Integration

This example includes an OpenAICompatibleClient which works with:

• OpenAI’s chat/completions
• DeepInfra, Fireworks, OpenRouter, Groq (any OpenAI-compatible proxy)

It supports both:

• Chat(ctx, messages) → full assistant reply
• StreamChat(ctx, messages) → streaming response via SSE

The orchestrator wraps these and exposes both:

HandleUserInput(...)        // streaming (returns StreamResult)
HandleUserInputFull(...)    // full response

---

🧪 Testing & Extending

Want to add a custom LLM or summarizer? Just implement:

• llm.Client interface for chat/stream
• prompt.FactSummarizer for memory compression

Then wire them into the orchestrator like this:

orchestrator := chat.NewDefaultOrchestrator(sessionMgr, promptBuilder, llmClient)

---

🧹 Cleanup

docker stop dragonfly && docker rm dragonfly

---

📜 License

MIT – see LICENSE