Awesome AI Agents › AI Assistants

vakovalskii/sgr-deep-research

⭐ 1118 Python repository created 2025-08-26

SGR Deep Research is the deep-research application layer of the SGR Agent Core project (the repository now redirects to vamplabAI/sgr-agent-core). It implements Schema-Guided Reasoning, an approach that constrains an LLM to emit structured output describing its next reasoning step and tool choice, so that a research loop stays inspectable instead of relying on free-form chain-of-thought. The framework exposes a BaseAgent interface with a two-phase architecture and ships several ready-made agent types: SGRAgent, ToolCallingAgent and a hybrid SGRToolCallingAgent. Around the agents sits a set of extensible tools for web search, page content extraction, reasoning steps and clarification requests, so an agent can ask the user a follow-up question before continuing a search. Deployment is configuration-driven: agents, LLM endpoints and tool credentials are declared in a config.yaml file, and the same file drives every entry point. The project runs as an HTTP service exposing OpenAI-compatible endpoints with server-sent-event streaming, which lets existing OpenAI clients point at it without code changes, and Swagger documentation is served alongside. A container image is published to GHCR, and the library itself installs from PyPI as sgr-agent-core for embedding in other Python code. Two command-line tools accompany the server: sgr starts the API server, and sgrsh gives an interactive shell with single-query and chat modes that handles clarification and intermediate-result requests. An sgracp binary speaks the Agent Client Protocol over stdio using newline-delimited JSON-RPC, so editors that support ACP can drive the agents directly. Because it targets any OpenAI-compatible endpoint, it also works with locally hosted models for private research. The authors report 86.08 percent accuracy on the SimpleQA benchmark with gpt-4.1-mini. It suits developers building research or question-answering agents who want structured, reproducible reasoning rather than opaque prompting.

https://github.com/vakovalskii/sgr-deep-research

pythondeep-researchschema-guided-reasoningstructured-outputfunction-callingopenai-apiagent-frameworkstreamingcliacp