Skip to main content

Using Knowledge

The knowledge tool itself has two modes of operation: Standalone and Server Mode - Check the sections below to learn more about them.

Both modes are configured the same way, via environment variables or command line flags:

Configuration​

Embedding Model Provider (must have)​

The model provider is the provider of the embeddings model that is used to encode ingested documents. Currently, we only support OpenAI and Azure OpenAI via the following flags / environment variables:

--openai-api-base string           OpenAI API base ($OPENAI_BASE_URL) (default "https://api.openai.com/v1")
--openai-api-key string            OpenAI API key ($OPENAI_API_KEY) (default "sk-foo")
--openai-api-type string           OpenAI API type (OPEN_AI, AZURE, AZURE_AD) ($OPENAI_API_TYPE) (default "OPEN_AI")
--openai-api-version string        OpenAI API version (for Azure) ($OPENAI_API_VERSION) (default "2024-02-01")
--openai-azure-deployment string   Azure OpenAI deployment name (overrides openai-embedding-model, if set) ($OPENAI_AZURE_DEPLOYMENT)
--openai-embedding-model string    OpenAI Embedding model ($OPENAI_EMBEDDING_MODEL) (default "text-embedding-ada-002")

Those are persistent flags, so they can be set on any knowledge subcommand.

1. Standalone Mode (Default)​

In standalone mode, Knowledge makes use of an embedded database and embedded Vector Database which the client connects to directly. This is the default and most simple mode of operation and is useful for local usage and offers a great integration with GPTScript.

Run the Client​

Any knowledge command (except for knowledge server) will use the standalone client mode, if no KNOW_SERVER_URL environment variable is set.

2. Server Mode​

In server mode, Knowledge uses a separate server for the Vector Database and the Document Database. This mode is useful when you want to share the data with multiple clients or when you want to use a more powerful server for the Vector Database.

Run the Server​

knowledge server

Ingestion​

To ingest a document, you can use the knowledge ingest command:

knowledge ingest --dataset my-dataset ./path/to/my-document.txt
note

By default, the dataset will be created if it doesn't exist. If you don't want that, you can use the --no-create-dataset flag.

Ignoring Files​

You can ignore files by providing an ignore file, similar to .gitignore:

knowledge ingest --dataset my-dataset --ignore-file .knowledgeignore ./path/to/my-documents

Here's an example ignore file which basically tells knowledge to only consider Markdown files and nothing else:

# Ignore everything
*

# Except Markdown files in any directory
!**/*.md
note

Alternatively, you can use the --ignore-extensions flag to ignore files with specific extensions.

knowledge ingest --dataset my-dataset --ignore-extensions=.txt ./path/to/my-documents

Remote Files​

You can ingest remote files by providing a URL - Currently only Git Repositories are supported:

knowledge ingest --dataset my-dataset https://github.com/gptscript-ai/knowledge

You may also specify a Commit Hash, Tag or Branch that you want to check out:

knowledge ingest --dataset my-dataset https://github.com/gptscript-ai/knowledge@v0.3.0
note

Here, it's advisable to use a ignore file to avoid ingesting all the git metadata and potentially present vendor files.