Open-source AI co-scientist style idea generation
Saluki turns research questions into ranked hypotheses.
Give Saluki a question and it generates ideas, critiques assumptions, scores them, runs tournament-style comparisons, evolves stronger variants, and prints the active research directions.
Idea generation needs selection pressure.
Saluki is built for the point where one-shot brainstorming stops being useful: when ideas need evidence, criticism, ranking, and evolution.
First ideas are cheap
A single pass can produce plausible hypotheses, but it rarely tells you which assumptions are weak or which ideas deserve more compute.
Ranking is hard
Novelty, rigor, testability, impact, and simplicity pull in different directions unless the comparison process is explicit.
Exploration stalls
Idea generation often collapses into nearby variants. Saluki keeps generating, critiquing, ranking, pruning, and evolving.
From prompt to evolving research portfolio
Saluki turns the research question into a task queue, then keeps cycling through generation, verification, comparison, and evolution.
Parse
Turn the research question into an objective, sub-questions, constraints, and search keywords.
Research
Optionally pull web evidence with Exa or Tavily before generating candidates.
Generate
Create hypotheses with a mechanism, prediction, and simple experiment.
Critique
Break ideas into assumptions, check plausibility, and reflect on weaknesses.
Rank
Score ideas across five dimensions and run pairwise tournament comparisons.
Evolve
Combine, simplify, or diverge from strong contenders to make better variants.
What Saluki generates
Ranked hypotheses with testable structure.
Every active idea is shaped around a mechanism, a prediction, and an experiment, then annotated with scores that make tradeoffs visible.
Hypothesis h0_0
Scores — novelty: 0.95 rigor: 0.85 testability: 0.70 impact: 0.40 simplicity: 0.30
Mechanism: A specific measurable mechanism explains the observed behavior.
Prediction: If the mechanism is real, a concrete measurable pattern should appear.
Experiment: Run the simplest validation that can separate the hypothesis from a plausible alternative.
Built for longer ideation runs
Task-queue supervisor
A supervisor schedules independent idea work, runs batches in parallel when allowed, and decides the next refinement step as the queue drains.
Tournament pressure
Ideas enter an Elo-style tournament, with similarity checks that avoid wasting matches on near-duplicates.
Deep reflection
Reflection and verification nodes expose weak assumptions before an idea rises through the standings.
Pause and resume
State persists under ~/.saluki/states by default, so longer ideation runs can be stopped and continued later.
Run the mode that matches the question
Scientific hypotheses
Use Saluki for questions where mechanisms, predictions, and experiments matter.
uv run saluki --prompt "What are promising biomarkers for early sepsis detection?"Decision framing
Ask for actions, target users, and reversal conditions when an idea should become a decision candidate.
uv run saluki --prompt "Where should we focus next?" --decision-modeNegative space
Make each idea state what it will not do, so scope boundaries are visible during comparison.
uv run saluki --prompt "New product research direction" --negative-spaceInstall Saluki
Clone the repo, sync the Python environment with uv, set an OpenAI-compatible API key, and run Saluki from the command line.
git clone https://github.com/20minds/saluki-idea-generation.git
cd saluki-idea-generation
uv sync
uv run saluki --prompt "What are promising biomarkers for early sepsis detection?"Frequently Asked Questions
Is Saluki a hosted co-scientist product?
No. Saluki is an open-source command-line tool that runs against OpenAI-compatible models and optional web-search providers.
Does Saluki need web search?
No. Saluki can run with only an OpenAI-compatible API key. If EXA_API_KEY or TAVILY_API_KEY is configured, research and reflection steps can use search results.
What does the tournament rank?
Each hypothesis is scored for novelty, rigor, testability, impact, and simplicity, then compared in pairwise matches so stronger ideas can rise.
Can I pause long runs?
Yes. Use --pause to persist state and --continue to resume the latest run. Saluki stores states under ~/.saluki/states unless you pass --persist-dir.
Want to use Saluki for research idea generation?