Knowledge Base

naaf-knowledge integrates Qdrant-backed retrieval with naaf-llm without making naaf-llm itself knowledge-aware.

The recommended pattern is:

Persist and select KnowledgeGroups in application code.
Build a KnowledgeLlmSession from those groups.
Reuse the generated system prompt and tool registry across requests.

Recommended Pattern

#![allow(unused)]
fn main() {
use naaf_knowledge::{KnowledgeGroup, KnowledgeLlmSessionBuilder};
use naaf_llm::{Executor, ExecutorConfig, OpenAiClient, OpenAiConfig};

let qdrant = naaf_qdrant::QdrantClient::from_url(
    "http://localhost:6333",
    Option::<String>::None,
)?
.with_collection("docs");
let embedder = naaf_qdrant::OpenAiEmbedder::new(std::env::var("OPENAI_API_KEY")?);

let knowledge = KnowledgeLlmSessionBuilder::new(Box::new(embedder))
    .with_system_prompt("You are a helpful assistant for the workspace.")
    .with_group(
        KnowledgeGroup::new("docs", "Documentation", "Product and API documentation"),
        qdrant,
    )
    .with_search_defaults(5, 0.7)
    .with_lint_tool(true)
    .build()?;

let llm_client = OpenAiClient::new(OpenAiConfig::new(std::env::var("OPENAI_API_KEY")?));
let executor = Executor::with_tools(llm_client, knowledge.tools().clone())
    .with_config(ExecutorConfig::new(5));

let request = knowledge.request_with_user_message("gpt-4o", "How do steps retry?");
let outcome = executor.execute(&(), request).await?;
println!("{}", outcome.final_message().content.as_deref().unwrap_or(""));
Ok::<(), Box<dyn std::error::Error>>(())
}

Why This Shape

naaf-llm stays generic and only owns execution and tool calling.
naaf-knowledge owns knowledge groups, prompt augmentation, and knowledge-specific tools.
The application chooses which groups are exposed for each request.

Main Types

KnowledgeGroup: metadata describing one exposed collection.
KnowledgeLlmSessionBuilder: high-level helper that creates the prompt and tools together.
KnowledgeLlmSession: reusable session object for request building.
KnowledgeSearchTool: lower-level retrieval tool for custom wiring.
KnowledgeLintTool: lower-level lint tool for custom wiring.

Prompt Augmentation

If you need more control than KnowledgeLlmSessionBuilder provides, naaf-knowledge also exposes prompt primitives:

#![allow(unused)]
fn main() {
use naaf_knowledge::{
    KnowledgePromptConfig, augment_system_prompt, format_knowledge_prompt_block,
};

let prompt_config = KnowledgePromptConfig::default();
let system_prompt = augment_system_prompt(
    "You are a helpful assistant.",
    &groups,
    &prompt_config,
);

let block = format_knowledge_prompt_block(&groups, &prompt_config);
}

When To Drop Down A Level

Use KnowledgeSearchTool and KnowledgeLintTool directly when:

You already have your own prompt-building pipeline.
You want custom CompletionRequest assembly.
You want to expose only search or only lint.

Keyboard shortcuts