An agentic research orchestrator for Go, Laconic addresses the challenge of prompt overflow in LLM agents by implementing state compression techniques. This makes agents practical for free search and low-cost, small-context window LLMs (4k/8k tokens). It benefits users by enabling cost-effective, efficient LLM-powered research and complex query answering.

How It Works

Laconic offers two core strategies for state compression and knowledge accumulation. The Scratchpad strategy employs a linear loop: Planner -> Search -> Synthesizer -> Finalizer. It compresses state by overwriting a free-text knowledge summary at each iteration, bounding prompt size. This is ideal for simple questions and minimal LLM calls. The Graph Reader strategy builds a notebook of atomic facts by traversing a dynamically constructed graph of search queries. It supports multi-hop reasoning and deep page reading via an optional FetchProvider, making it suitable for complex research where answer quality is paramount, though it incurs higher LLM costs.

Quick Start & Requirements

• Installation: go get github.com/smhanov/laconic
• Prerequisites: Go development environment. Requires implementing LLMProvider adapters (e.g., for OpenAI, Ollama, Anthropic, Gemini via llmhub) and configuring SearchProvider (DuckDuckGo, Brave, Tavily). An optional FetchProvider can be integrated for full-page reading.
• Resources: Minimal dependencies (stdlib only). Example implementations are available in examples/basic/ and examples/research/.
• Documentation: Detailed architecture and prompt design are available in docs/architecture.md and docs/prompts.md.

Highlighted Details

• Dual Strategies: Scratchpad for efficiency and Graph Reader for depth.
• Cost Management: Automatic LLM and search cost tracking.
• Knowledge Persistence: Supports follow-up questions via Result.Knowledge carry-over.
• Model Agnosticism: Integrates with any LLM or search provider via interfaces.
• Dual-Model Support: Allows distinct models for planning and synthesis/finalization.
• Pluggable System: Custom strategies can be registered.

Maintenance & Community

No specific details regarding maintainers, community channels (e.g., Discord, Slack), or active sponsorships were found in the provided README.

Licensing & Compatibility

• License: MIT.
• Compatibility: The MIT license is highly permissive, allowing for commercial use and integration into closed-source projects without significant restrictions.

Limitations & Caveats

The Graph Reader strategy is more LLM-intensive and benefits from larger context windows (16k+ tokens recommended), whereas Scratchpad is optimized for smaller contexts (4k-8k tokens) and fewer LLM calls. The FetchProvider is optional but crucial for the Graph Reader's deep-reading capabilities.