CLI

Both packages install a raglite command with the same commands and flags.

TypeScript

npm install -g raglite-toolkit @huggingface/transformers
# or run without installing:
npx -y -p raglite-toolkit -p @huggingface/transformers raglite --help

Python

pip install 'raglite-toolkit[local]'
raglite --help

Every command takes a <path|url>: a file, a folder (indexed as a collection) or a web URL. The index is reused between commands, so search after index embeds only the query.

raglite index ./docs
raglite search ./docs "refund policy" --top-k 3
raglite search ./docs "ERR_4021" --mode hybrid
raglite ask ./docs "What is the refund policy?" --llm-provider anthropic --stream
raglite serve ./docs --llm-provider openai --port 8085 --token "$RAGLITE_TOKEN"
raglite mcp ./docs --store-dir "$PWD/.raglite"
raglite eval ./docs questions.jsonl --rerank cohere

Commands

index

Builds or reuses the index and prints the build result as JSON.

Flag	Meaning
--chunk-size N	Words per chunk (characters for Chinese, Japanese, Thai, …)
--overlap N	Overlap between chunks
--rebuild	Re-index even if nothing changed

search

raglite search <path|url> "query" prints the results as JSON.

Flag	Meaning
--top-k N	Number of results
--mode M	vector, keyword or hybrid

ask

raglite ask <path|url> "question" prints the answer, then a Sources: list of the cited passages with page and section.

Flag	Meaning
--llm-provider P	Required. openai, anthropic, google, mistral, cohere, groq, xai, ollama
--llm-model M	Model; default from Choosing providers
--llm-key K	API key; default from the provider’s environment variable
--top-k N	Passages given to the LLM
--mode M	Retrieval mode
--stream	Print the answer as it is written (no Sources: list)

serve

Starts the REST API and runs until stopped.

Flag	Meaning
--llm-provider P, --llm-model M, --llm-key K	LLM for /ask; without one, /ask returns 503
--host H	Default 127.0.0.1
--port N	Default 8085
--token T	Bearer token required on every endpoint except /health
--mode M	Default retrieval mode

mcp

Serves the MCP server over stdio. Takes --llm-provider, --llm-model and --llm-key (which add the ask tool) and --mode. The TypeScript package needs @modelcontextprotocol/sdk; the Python package needs the [mcp] extra.

eval

raglite eval <path|url> questions.jsonl scores retrieval on your questions: recall@k, MRR and hit rate for each mode. See Evaluating retrieval.

Flag	Meaning
--top-k N	Results scored per question (default 5)
--modes M,M	Modes to compare, comma-separated (default vector,keyword,hybrid)
--rerank P	Also score each mode reranked by cohere, voyage or llm
--rerank-model M, --rerank-key K	Reranking model and API key
--llm-provider P, --llm-model M, --llm-key K	LLM for --rerank llm
--json	Print the report as JSON

Flags on every command

Flag	Meaning
--store-dir D	Where indexes are kept (default ./.raglite)
--embed-provider P	Embedding provider; default: the existing index’s, or local for a new one
--embed-model M	Embedding model
--embed-key K	Embedding API key

Use the same --store-dir and embedding flags for every command on the same index.

Vector store flags (TypeScript only)

Flag	Meaning
--vector-provider P	memory, qdrant, pinecone or lancedb
--vector-url U	Qdrant URL or Pinecone index host
--vector-key K	API key for the store
--vector-index I	Shared Qdrant collection name
--vector-store-dir D	Folder for the memory and LanceDB stores

Output and scripting

Results go to stdout and logs to stderr, so JSON output can be piped:

raglite search ./docs "refund policy" --top-k 3 | jq -r '.[] | "\(.score)  \(.metadata.source)"'

raglite --version prints the version.