> ## Documentation Index
> Fetch the complete documentation index at: https://praison.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Quality-Based RAG

> Score, filter, and rerank retrieved content for higher-quality answers

Retrieve with quality scoring — filter low-quality chunks and rerank before the agent answers.

```python theme={"theme":{"light":"vitesse-light","dark":"vitesse-dark"}}
from praisonaiagents import Agent, KnowledgeConfig

agent = Agent(
    name="Quality RAG",
    instructions="Answer from documents. Prefer high-quality sources.",
    knowledge=KnowledgeConfig(
        sources=["docs/"],
        rerank=True,
        retrieval_k=5,
        retrieval_threshold=0.7,
    ),
)

agent.start("What are the key product features?")
```

The user asks a question; retrieval scores and filters chunks before the agent answers from high-quality sources.

```mermaid theme={"theme":{"light":"vitesse-light","dark":"vitesse-dark"}}
graph LR
    Q[Query] --> R[Retrieve]
    R --> S[Quality score]
    S --> F[Filter]
    F --> K[Rerank]
    K --> A[Answer]

    classDef query fill:#8B0000,stroke:#7C90A0,color:#fff
    classDef process fill:#189AB4,stroke:#7C90A0,color:#fff
    classDef output fill:#10B981,stroke:#7C90A0,color:#fff

    class Q query
    class R,S,F,K process
    class A output
```

## How It Works

```mermaid theme={"theme":{"light":"vitesse-light","dark":"vitesse-dark"}}
sequenceDiagram
    participant User
    participant Agent
    participant QualityBasedRag

    User->>Agent: Request
    Agent->>QualityBasedRag: Process
    QualityBasedRag-->>Agent: Result
    Agent-->>User: Response
```

## Quick Start

<Steps>
  <Step title="Simple Usage">
    Enable reranking on the agent's knowledge base:

    ```python theme={"theme":{"light":"vitesse-light","dark":"vitesse-dark"}}
    from praisonaiagents import Agent, KnowledgeConfig

    agent = Agent(
        name="RAG Agent",
        instructions="Answer from the knowledge base.",
        knowledge=KnowledgeConfig(sources=["docs/"], rerank=True),
    )

    agent.start("Summarise the onboarding guide")
    ```
  </Step>

  <Step title="With Configuration">
    Filter memory search by quality score:

    ```python theme={"theme":{"light":"vitesse-light","dark":"vitesse-dark"}}
    from praisonaiagents import Agent, Memory

    agent = Agent(
        name="Memory RAG",
        instructions="Use only high-quality past outputs.",
        memory=Memory(),
    )

    results = agent.memory.search_short_term(
        "previous research on AI",
        min_quality=0.7,
        rerank=True,
        limit=5,
    )
    ```
  </Step>
</Steps>

***

## How It Works

```mermaid theme={"theme":{"light":"vitesse-light","dark":"vitesse-dark"}}
sequenceDiagram
    participant User
    participant Agent
    participant Feature as Quality-Based RAG

    User->>Agent: Request
    Agent->>Feature: Process request
    Feature-->>Agent: Result
    Agent-->>User: Response
```

Quality spans retrieval and memory:

| Layer             | Mechanism                                                                     |
| ----------------- | ----------------------------------------------------------------------------- |
| **Knowledge**     | `KnowledgeConfig(rerank=True)` reranks chunks before injection                |
| **Memory search** | `min_quality` filters by stored quality metadata                              |
| **Memory store**  | `compute_quality_score()` combines completeness, relevance, clarity, accuracy |
| **Task output**   | `quality_check=True` on tasks stores high-scoring outputs (threshold 0.7)     |

```python theme={"theme":{"light":"vitesse-light","dark":"vitesse-dark"}}
score = memory.compute_quality_score(
    completeness=0.9,
    relevance=0.85,
    clarity=0.8,
    accuracy=0.9,
)
```

***

## Configuration Options

### KnowledgeConfig (retrieval)

| Option                | Type    | Default | Description              |
| --------------------- | ------- | ------- | ------------------------ |
| `rerank`              | `bool`  | `False` | Rerank retrieved chunks  |
| `rerank_model`        | `str`   | `None`  | Model for reranking      |
| `retrieval_k`         | `int`   | `5`     | Chunks to retrieve       |
| `retrieval_threshold` | `float` | `0.0`   | Minimum similarity score |

### Memory search

| Parameter          | Type    | Default | Description              |
| ------------------ | ------- | ------- | ------------------------ |
| `min_quality`      | `float` | `0.0`   | Minimum quality metadata |
| `rerank`           | `bool`  | `False` | Rerank search results    |
| `relevance_cutoff` | `float` | `0.0`   | Minimum relevance score  |

***

## Best Practices

<AccordionGroup>
  <Accordion title="Enable rerank for accuracy-critical tasks">
    Reranking adds an LLM call — disable for high-throughput pipelines.
  </Accordion>

  <Accordion title="Set retrieval_threshold for noisy corpora">
    Raise `retrieval_threshold` to 0.6–0.7 when documents overlap heavily.
  </Accordion>

  <Accordion title="Combine with task quality_check">
    Use `Task(quality_check=True)` so only high-quality outputs enter long-term memory.
  </Accordion>

  <Accordion title="Weight metrics for your domain">
    Pass custom `weights` to `compute_quality_score()` — weight accuracy higher for factual Q\&A.
  </Accordion>
</AccordionGroup>

***

## Related

<CardGroup cols={2}>
  <Card title="Quality Checking" icon="check-circle" href="/docs/features/quality-checking">
    Automatic task output assessment
  </Card>

  <Card title="Knowledge" icon="book" href="/docs/features/knowledge">
    Knowledge base setup and retrieval
  </Card>
</CardGroup>