A3S CodeExamples
Vector RAG
Semantic code search via vector embeddings
Vector RAG
A3S Code includes a VectorContextProvider that indexes workspace files and retrieves semantically relevant code snippets using cosine similarity. Unlike grep (exact pattern matching), this finds code by meaning.
Vector RAG requires an embedding API. Configure embedding_api_base and embedding_api_key in your provider config, or use the codesearch tool which auto-indexes on first query.
FileSystem Context (No Embedding Required)
The simplest context provider — injects relevant file contents based on keyword matching.
use a3s_code_core::{Agent, SessionOptions};
let opts = SessionOptions::new()
.with_permissive_policy()
.with_fs_context("/my-project"); // inject relevant files into context
let session = agent.session("/my-project", Some(opts))?;
let result = session.send("How does authentication work?", None).await?;
println!("{}", result.text);Run: cargo run --example test_vector_rag
Source: core/examples/test_vector_rag.rs
session = agent.session("/my-project",
permissive=True,
fs_context="/my-project",
)
result = await session.send("How does authentication work?")
print(result.text)const session = agent.session('/my-project', {
permissive: true,
fsContext: '/my-project',
});
const result = await session.send('How does authentication work?');
console.log(result.text);Vector Context Provider (Semantic Search)
use a3s_code_core::context::{VectorContextConfig, VectorContextProvider};
use a3s_code_core::context::embedding::OpenAiEmbeddingProvider;
use a3s_code_core::context::vector_store::InMemoryVectorStore;
use a3s_code_core::{Agent, SessionOptions};
use std::sync::Arc;
// Build the vector provider
let embedder = OpenAiEmbeddingProvider::new(
"https://api.openai.com/v1",
std::env::var("OPENAI_API_KEY")?,
"text-embedding-3-small",
1536,
);
let config = VectorContextConfig::new("/my-project")
.with_min_relevance(0.3)
.with_max_results(10);
let store = InMemoryVectorStore::new();
let provider = Arc::new(VectorContextProvider::new(config, embedder, store));
// Index the workspace
provider.index().await?;
// Wire into session
let opts = SessionOptions::new()
.with_permissive_policy()
.with_context_provider(provider);
let session = agent.session("/my-project", Some(opts))?;
let result = session.send("Find all JWT authentication code", None).await?;
println!("{}", result.text);# Vector context is configured via the agent config file.
# Add to ~/.a3s/config.hcl:
#
# context {
# provider = "vector"
# embedding_model = "text-embedding-3-small"
# min_relevance = 0.3
# }
session = agent.session("/my-project", permissive=True)
result = await session.send("Find all JWT authentication code")
print(result.text)// Vector context is configured via the agent config file.
// Add to ~/.a3s/config.hcl:
//
// context {
// provider = "vector"
// embedding_model = "text-embedding-3-small"
// min_relevance = 0.3
// }
const session = agent.session('/my-project', { permissive: true });
const result = await session.send('Find all JWT authentication code');
console.log(result.text);CodeSearch Tool
The codesearch tool exposes vector search as an LLM-callable tool. The LLM can invoke it directly when it needs to find code by meaning.
use a3s_code_core::tools::builtin::codesearch::create_codesearch_tool;
use a3s_code_core::context::embedding::OpenAiEmbeddingProvider;
// Create and register the codesearch tool
let embedder = OpenAiEmbeddingProvider::new(/* ... */);
let tool = create_codesearch_tool("/my-project", embedder);
let opts = SessionOptions::new().with_permissive_policy();
let session = agent.session("/my-project", Some(opts))?;
session.register_tool(Arc::new(tool));
// Now the LLM can call: codesearch(query="JWT authentication")
let result = session.send(
"Search for code related to JWT token verification",
None,
).await?;
println!("{}", result.text);# The codesearch tool is available when vector context is configured.
# The LLM will use it automatically when searching for code by meaning.
result = await session.send("Search for code related to JWT token verification")
print(result.text)// The codesearch tool is available when vector context is configured.
const result = await session.send('Search for code related to JWT token verification');
console.log(result.text);Chunking Strategy
Files are split on paragraph/function boundaries with a configurable max chunk size (default: 1500 chars). Each chunk is embedded independently and stored with its source file path and relevance score.
auth.rs → [chunk 1: verify_jwt fn] [chunk 2: Claims struct] [chunk 3: tests]
db.rs → [chunk 1: connect_pool fn] [chunk 2: Pool config]API Reference
SessionOptions
Prop
Type
VectorContextConfig (Rust)
Prop
Type
HCL config (Python / Node.js)
context {
provider = "vector"
embedding_model = "text-embedding-3-small"
min_relevance = 0.3
max_results = 10
}codesearch tool schema
Prop
Type