Collections

A DocumentCollection indexes many sources and searches them as one. Each source is a file, a folder (scanned recursively) or a web URL.

TypeScript
Python

import { DocumentCollection } from "raglite-toolkit";

const collection = new DocumentCollection(["./docs", "./faq.md", "https://example.com/pricing"], {
  llm: { provider: "openai" },
});
collection.addSource("./handbook.pdf");

const result = await collection.build();
console.log(`${result.totalDocuments} documents, ${result.totalChunks} chunks`);

const hits = await collection.search("refund policy", { topK: 5 });
const answer = await collection.ask("What is the refund policy?");

from raglite import DocumentCollection

collection = DocumentCollection(["./docs", "./faq.md", "https://example.com/pricing"], {
    "llm": {"provider": "openai"},
})
collection.add_source("./handbook.pdf")

result = collection.build()
print(f"{result.totalDocuments} documents, {result.totalChunks} chunks")

hits = collection.search("refund policy", top_k=5)
answer = collection.ask("What is the refund policy?")

Each file gets its own index, so build() re-embeds only the files that changed. Search results from every document are merged by score; in hybrid mode the lists are merged first, then fused once.

Folders

A folder is scanned recursively for supported files (PDF, DOCX, Markdown, TXT, JSON).

Unsupported and empty files are skipped and reported in errors.
Hidden files and folders (such as .git, .venv and .raglite) and node_modules are skipped.

Glob patterns such as ./docs/*.md are not expanded. Pass a folder, or a list of files.

The build result

Field	Meaning
`totalDocuments`	Documents in the collection
`totalChunks`	Chunks across all documents
`cachedDocuments`	Documents whose saved index was reused
`newDocuments`	Documents indexed in this build
`errors`	`{ source, error }` for each source that failed

One broken file does not stop the build: it is reported in errors and the rest are indexed. A search fails only when every document fails.

Vector stores

Collections work with every vector store. Pass a provider config such as { provider: "qdrant", ... }, not a VectorStore instance: an instance has one namespace, and every document would overwrite the others. RAGLite throws a ConfigError if you try.

Other methods

TypeScript	Python	Does
`askStream()`	`ask_stream()`	Streams the answer. See Streaming
`asTool()`	`as_tool()`	A search tool for agents. See Agent tools
`serve()`	`serve()`	A REST API. See REST API
`getDocuments()`	`get_documents()`	The `Document` for each source
`chunkCount`	`chunk_count`	Chunks across all documents