Pao Siangliulue, Jonathan Bragg, Doug Downey, Joseph Chee Chang, Daniel S. Weld
4 min
Omakase is an innovative AI research assistant designed to support scientists over the long haul of their projects. Unlike traditional AI tools that wait for you to ask a question, Omakase is proactive—it watches your project documents as they evolve, figures out what you might need next, queries powerful 'deep research systems' (advanced AI that synthesizes vast literature), and boils down their lengthy reports into actionable suggestions. Delivered via email notifications and a dashboard, these suggestions are tailored to your specific project context, helping you spot concrete next steps like "Read this paper on X method—it addresses the gap in your hypothesis on page 5."
This addresses two big pain points in research: (1) Manually crafting precise queries for complex, ongoing projects is tedious—you have to cram rich context into short prompts. (2) Deep research tools spit out exhaustive reports that are overwhelming to sift through for real actions. Omakase automates both, making AI a true partner in long-horizon scientific work.
Traditional AI interactions are reactive: single-turn queries or multi-turn chats. Proactive agents take initiative, but prior work focused on short sessions (e.g., clarification questions). Omakase pioneers long-horizon proactivity by monitoring evolving project artifacts like shared docs. It infers 'latent needs'—unspoken questions based on your project's state—and acts asynchronously via notifications, not real-time chat. This fits research workflows where needs shift slowly over weeks/months, unlike coding or design tasks.
Key enabler: Project context extraction from lightweight, privacy-friendly sources like a 'Research Interests Document' (your notes + seed papers), avoiding screenshots or chat histories.
Omakase queries deep research systems (e.g., LLM-powered tools like those in [49]) which excel at complex knowledge synthesis but produce verbose outputs. Omakase distills these into actionable suggestions:
The paper recommendations pipeline is LLM-driven:
Built through technology probe studies (N=42 total): formative studies (N=28), 10 pipeline iterations, final eval (N=11). Researchers rated inferred queries as useful/timely and Omakase suggestions significantly more actionable than raw deep research reports. Interfaces include a dashboard for suggestion lists (with doc context + QA excerpts) and document views highlighting relevant sections.
As LLMs handle complex tasks, the bottleneck shifts to context and proactivity. Omakase extends RAG-based literature tools (paper recs, synthesis) to whole-project awareness, where literature needs evolve. It previews a future of AI as 'ambient intelligence' for science—always aware, never intrusive—potentially accelerating discovery by surfacing serendipitous insights at the right time. Challenges remain: ensuring query accuracy, handling project ambiguity, and scaling to diverse fields.
As AI agents become increasingly capable of complex knowledge tasks, the lack of context limits their capability to proactively reason about a user's latent needs throughout a long evolving project. In scientific research, many researchers still manually query a deep research system and compress their rich project contexts into short, targeted queries. Further, a deep research system produces exhaustive reports, making it difficult to identify concrete actions. To explore the opportunities of research assistants that are proactive throughout a research project, we conducted several studies (N=42) with a technology probe and an iterative prototype. The latest iteration of our system, Omakase, is a research assistant that monitors a user's project documents to infer timely queries to a deep research system. Omakase then distills long reports into suggestions contextualized to their evolving projects. Our evaluations showed that participants found the generated queries to be useful and timely, and rated Omakase's suggestions as significantly more actionable than the original reports.
Alex: Even blind, people picked up-to-date ones as better fits.
Sam: Yes. For suggestions versus raw reports—where folks picked promising sections—Omakase's rated higher on concrete steps, like "adopt this baseline," even against reading twice as much raw text. Nine of ten found them easy to scan. The paper suggests distilling reports into project actions is a meaningful step.
Alex: That addresses drowning in text. Users could steer timing, like weekly checks.
Sam: Right—they set frequencies or track specific questions for updates. With eleven researchers, four of ten preferred Omakase's boiled-down ideas over raw reports. They felt like advisor feedback—pointing to overlooked experiment choices—and connected easily back to the project.
Alex: But were there misses?
Sam: Yes—jumping to assumptions about project direction led to off-target ideas. It sometimes pulled from old versions with abandoned topics. The paper notes this as room for improvement, like flagging dropped threads. Privacy came up too—hesitation due to rules or AI access worries. Studies were short-term, so long-term use over full projects needs more testing.
Alex: Balancing awareness with caution is key.
Sam: Shared docs give reliable, privacy-friendlier snapshots than chat logs or screenshots, which mix notes or fade. The paper cautions on risks like over-reliance without verifying.
Alex: Overall, this takes paper-retrieval tools and makes them smarter by tying into changing project documents.
Sam: Precisely. Iterative design with 42 participants refined it into precise, linked suggestions from evolving docs.
Alex: That's a clear takeaway—smarter support without the grind. Thanks, Sam. And thanks for listening to ResearchPod.