蚂蚁 inclusionAI 开源 PanelWise 多模型深度研究系统

inclusionAI; PanelWise

Repository navigation

NotificationsYou must be signed in to change notification settings; Fork0 (0); Star134 (134)You must be signed in to star a repository

Self-hosted multi-model deep research system that fuses independent research agents into evidence-grounded reports.

Repository files navigation

README; Contributing; Apache-2.0 license; Security

English · 中文

Fuse the complementary strengths of multiple models into one stronger result.

PanelWise starts with a simple question: if every model sees something worth keeping, can their complementary judgments produce a result stronger than any one answer? It is an embeddable Python aggregation engine that sends the same task to a configurable panel, preserves each model's distinct contribution, and coordinates those contributions through one of two execution topologies. The task itself is unrestricted; prompts, providers, and environment adapters define the domain.

Two modes, one interface

These are execution modes, not hard-coded task categories. Both accept any task string. Custom ChatClient and Executor implementations can connect the same engine to another provider or environment.

--eval and --no-eval override one YAML value, so applications can switch topology without changing code or duplicating configuration.

Install

PanelWise requires Python 3.10 or newer.

Install a pinned Git revision directly:

Replace main with a release tag or commit SHA when reproducibility matters. Contributors should use python -m pip install -e '.[dev]'.

Five-minute start

Create a documented configuration and validate it without making model calls:

The generated YAML contains provider, model-role, concurrency, timeout, workspace, and prompt settings. The same file drives both modes:

See the full panelwise.example.yaml and configuration reference.

Python library

PanelWiseResult has one stable contract across both modes. It always includes request_id, mode, status, output, usage, and errors. Eval mode additionally returns independent candidates and structured evaluation; no-eval mode returns the shared trajectory and executor artifacts, including a Git patch when available.

Individual panel failures are isolated. Invalid configuration raises a documented PanelWiseError subclass; it never terminates the host process with SystemExit. See the Python API and failure contract.

Results on DRACO

We evaluated PanelWise on all 100 tasks in DRACO, a deep-research benchmark spanning ten domains. Each task is scored against a weighted rubric covering factual accuracy, breadth, depth, presentation, and citation quality.

The OpenRouter Fusion score of 68.3 and Claude Fable 5 score of 65.3 are both quoted from OpenRouter's official Fusion launch post. PanelWise scores come from our complete 100-task evaluation runs.

The frontier panel scored 5.38 points above the published Fusion result using the same three participant models. The independent budget run scored 1.12 points above the strongest published single-model baseline. OpenRouter reports Fable 5 over 93 completed tasks because content filters blocked seven tasks, so that comparison is slightly uneven.

The benchmark supports a claim about deep-research aggregation on DRACO; it is not a claim that every task or model panel improves.

Providers and extension points

The built-in async client supports:

OpenRouter; ZenMux; any OpenAI-compatible chat-completions base URL

Provider-specific request fields can be supplied through YAML. A custom provider implements the small ChatClient protocol. A custom shared environment implements Executor, allowing no-eval mode to coordinate a shell, container, browser, database, simulator, or remote worker without changing the aggregation engine.

Trajectory mode's built-in local shell executor rejects obvious destructive commands, enforces time and output bounds, and starts every action in one configured workspace. It is a guardrail, not a sandbox; untrusted tasks belong in a disposable container or VM.

Documentation

CLI reference; Configuration and provider setup; Python API and result contract; Examples; Troubleshooting; Release process; Contributing and pull request requirements; Security policy; Changelog

Development

The test suite is offline and requires no model key. Public API, provider, executor, and architectural changes should begin with an issue; focused documentation fixes may go directly to a PR. Every behavior change must include reproducible verification. See CONTRIBUTING.md.

Contributors

Contributors are displayed in the order defined in CONTRIBUTORS.md.

License

PanelWise is available under the Apache License 2.0.

Self-hosted multi-model deep research system that fuses independent research agents into evidence-grounded reports.