Specialized Solver Plugins

Solver plugins also exist for specialized tasks, like regular question solvers these also benefit from automatic bidirectional translation for language support

---

ReRankers / MultipleChoiceQuestionSolvers

A specialized kind of solver plugin that chooses the best answer out of several options

These specialized solvers are used internally by ovos-common-query-pipeline-plugin, some skills and even by other question solver plugins!

Example configuration of ovos-flashrank-reranker-plugin for usage with ovos-common-query-pipeline-plugin

"intents": {
    "common_query": {
        "min_self_confidence": 0.5,
        "min_reranker_score": 0.5,
        "reranker": "ovos-flashrank-reranker-plugin",
        "ovos-flashrank-reranker-plugin": {
          "model": "ms-marco-TinyBERT-L-2-v2"
        }
    }
}

---

Evidence Solver

Evidence solvers accept not only a question but also a companion piece of text containing the answer.

Some question solver plugins like ovos-solver-wikipedia-plugin use evidence solvers internally, they are often helpful to generate a question out of a search result

---

Summarizer

Some question solver plugin use summarizers internally, they are often helpful to shorten long text from web search results

---

Collaborative Agents via MoS (Mixture of Solvers)

One of the most powerful features of the OVOS solver architecture is its ability to orchestrate multiple agents collaboratively through specialized Mixture of Solvers (MoS) plugins.

These MoS solvers implement strategies that combine the strengths of various LLMs, rerankers, rule-based solvers, or even remote agents (like HiveMind nodes), allowing dynamic delegation and refinement of answers.

🤝 Flexible Plugin Design: MoS strategies are implemented as standard solver plugins. This means they can be composed, nested, or swapped just like any other solver—allowing advanced collaborative behavior with minimal integration effort.

How It Works

Instead of relying on a single model or backend, a MoS solver delegates the query to several specialized solvers (workers) and uses strategies like voting, reranking, or even further generation to decide the best final response.

Examples include:

• The King: Uses a central "king" (reranker or LLM) to select or generate the best answer based on multiple solver outputs.

---

• Democracy: Implements a voting system among reranker solvers to choose the most agreed-upon response.

---

• Duopoly: A pair of collaborating LLMs generate and discuss answers before passing them to a final decider ("the president" solver).

---

Each strategy enables different dynamics between solvers—be it a single judge, a voting panel, or a back-and-forth discussion between agents.

🌀 Recursive Composition: Any MoS strategy can recursively use another MoS as a sub-solver, allowing for arbitrarily deep collaboration trees.