Most web-research assistants either depend on cloud LLMs or produce a single-pass summary that misses open questions. This project flips that model by running an iterative, local-first research loop: a local LLM generates search queries, summarizes findings, reflects to identify gaps, then issues new queries to close those gaps—producing a reproducible research trace and a final markdown report with sources.
What Sets It Apart
- Local-first, model-agnostic execution: Designed to run with local LLMs served through Ollama or LMStudio, so sensitive queries can stay on-prem and researchers keep full control of the model runtime. (So what: reduces data exposure and eases reproducibility for offline or regulated environments.)
- Iterative search → summarize → reflect loop: Instead of one-shot scraping, it explicitly reflects on summaries to generate targeted follow-up searches and repeats for a configurable number of cycles. (So what: uncovers knowledge gaps and produces more thorough, multi-pass synthesis.)
- Pluggable search backends and visualization: Defaults to DuckDuckGo but supports other search APIs; integrates with a LangGraph Studio UI to visualize collected sources and the final summary. (So what: makes the research process auditable and the source set inspectable.)
- Practical fallbacks for model quirks: Includes mechanisms to switch from strict JSON-mode outputs to tool-calling when some local models cannot produce structured responses. (So what: improves robustness across a wide range of open/local models.)
Who it's for — Great fit if / Look elsewhere if
- Great fit if you need privacy-conscious, reproducible web research pipelines that run without cloud LLMs, or if you want an auditable multi-pass synthesis for topic exploration. Also useful for developers who want an off-the-shelf graph to visualize research state.
- Look elsewhere if you need a turnkey, cloud-hosted consumer search assistant with continuously updated web crawling and ranked answers, or if you cannot run a local LLM (Ollama/LMStudio) in your environment.
Where It Fits
This project sits between single-query summarizers and heavy-weight RAG platforms: it does not provide large-scale vector stores or managed crawling out of the box, but it automates iterative web retrieval and synthesis using local models, making it a sensible choice for privacy-aware teams and developers prototyping local research agents. Expect trade-offs: better control and auditability in exchange for requiring local model infrastructure and occasional model-specific configuration.
