GPTKB Studio

GPTKB Studio is a web application for building small, topic-specific knowledge bases from what a large language model knows. You choose a topic, a seed entity and a model; GPTKB Studio asks the model for facts about the seed as subject–predicate–object triples, recognizes the named entities among the objects, decides whether each one is an entity it already has or a new one, and then asks about the new ones, round after round, until nothing new is left or a budget runs out.

Entities, relations and classes are disambiguated during construction with the pipeline of GPTKB 2.0, so the result has canonical identifiers rather than strings, and every merge decision can be inspected.

A hosted instance runs at https://gptkbstudio.scads.ai/.

Features

• Any model, any OpenAI-compatible endpoint: presets for ScaDS.AI, OpenAI and OpenRouter, or a custom URL (e.g. a self-hosted vLLM server; on a local network only with MINIGPTKB_ALLOW_PRIVATE_ENDPOINTS=1). One model does all three tasks (fact elicitation, named entity recognition, disambiguation) unless NER and NED are given models of their own under Advanced.
• Topic-scoped, multilingual elicitation: 11 pre-packaged prompt languages, an editable prompt, open, example-driven or fixed-ontology schema.
• Budgets: caps on entities, BFS depth, cost (USD) and runtime. Entity and depth caps stop the exploration but let pending disambiguation finish.
• Live view: metrics, the current pipeline step and the log stream into the browser; runs can be paused, stopped and resumed, also after a server restart.
• Controllable merging: a slider trades fewer wrong merges for more duplicates by restricting which existing entities are offered as matches.
• Results: a browsable HTML report per run (statistics, an interactive graph, one page per entity with its disambiguation trace) and a SQLite database with entities, relations, classes, aliases, triples and every disambiguation decision.
• Shared, rate-limited endpoints: requests are paced per model below the limits the endpoint reports.

How a run works

Each run is a separate worker process. In every round it runs the GPTKB 2.0 stages on whatever items are ready for them:

NEDG -> NED2 -> NED1 -> Elicitation -> PDG -> CDG -> PD -> CD -> NER

• Elicitation asks for the facts about unexplored entities.
• NER classifies each triple's object as a named entity or a literal.
• NED1 matches an object against the most similar existing entities, using its triple as context; NEDG writes a description for objects without a match, and NED2 matches again using that description, or creates a new entity.
• PD/PDG and CD/CDG do the same for relations and classes.

Candidates are retrieved by label embedding (Qwen3-Embedding-4B). The state of a run lives in runs/<run id>/knowledgeBase.db, which is also its export and what a resumed run continues from.

The original crawler, which identifies entities by their names as in the miniGPTKB method it grew out of, is kept in app/runner.py; runs made with it stay viewable and resumable, and new ones can be started through the API with "pipeline": "classic".

Running it

Requirements: Python 3.11 or newer.

pip install -r requirements.txt
python -m uvicorn app.server:app --host 127.0.0.1 --port 8000

Then open http://127.0.0.1:8000.

Embeddings need a ScaDS.AI key. Disambiguation retrieves its candidates with the embedding model Qwen3-Embedding-4B, which GPTKB Studio calls on the ScaDS.AI inference service with the key in SCADS_API_KEY, whatever endpoint the language models use. Without that key, runs with disambiguation fail. To run GPTKB Studio without one, point the embeddings to another OpenAI-compatible embeddings endpoint (e.g. Qwen3-Embedding-4B served locally with vLLM or Ollama): the endpoint and model are set in app/disambig/runner.py (SCADS_EMBEDDING_URL, DEFAULT_EMBEDDING_MODEL) and the key in app/server.py. With a different embedding model, the similarity thresholds of the entity-merging slider and of label clustering may need adjusting.

Environment variables:

Variable	Purpose
SCADS_API_KEY	Key for the ScaDS.AI inference service. Used for label embeddings, and for model calls on ScaDS.AI when the user leaves the API key field empty.
MINIGPTKB_RUNS_DIR	Where run folders are kept (default: runs/).
MINIGPTKB_SHARED_KEY_MINUTES	How many minutes one run may use the shared key, in total over all its sessions; the server ends it after that (default: no limit).
MINIGPTKB_MAX_SHARED_KEY_RUNS	How many runs may use the shared key at the same time (default: no limit).
MINIGPTKB_SITE_LABEL	A label shown in red under the app name, e.g. Sandbox on an internal test instance (default: none).
MINIGPTKB_ALLOW_PRIVATE_ENDPOINTS	Set to 1 to allow model endpoints on private, loopback or link-local addresses (e.g. a model served with vLLM or Ollama on the same machine). By default such endpoints are refused, so that users of a hosted instance cannot make the server call its own network.

API keys that users enter in the form are passed to the run in memory only: they are never written to disk, and they are sent only to the endpoint they were entered for.

deploy/ contains an example systemd unit and environment file for a hosted instance.

A folder can be made read-only, e.g. for showcase runs on a public instance, by adding "locked": true to its entry in <runs dir>/_folders.json on the server. Nobody can then rename or delete it, delete, move, resume or stop its runs, or put other runs into it; its runs can still be explored, watched and downloaded.

Repository layout

Path	Contents
app/server.py	FastAPI server: REST API, run processes, event stream, run list and folders, downloads
app/disambig/	The disambiguation pipeline: runner, prompts and answer parsers, SQLite schema, embeddings, pacing, HTML report
app/runner.py	The original string-identity crawler
app/usage.py	Per-run counts of model calls and tokens (usage.json)
app/static/	The browser application (plain JavaScript, no build step)
app/vendor/	vis-network, for the graph view
tests/	Unit tests: python -m unittest discover -s tests

License

GPTKB Studio is released under the Creative Commons Attribution 4.0 International license.

The prompts in app/disambig/prompts/ are taken from the GPTKB 2.0 repository and remain under the Apache License 2.0 (app/disambig/prompts/LICENSE-GPTKB2, source commit in app/disambig/prompts/SOURCE). app/vendor/vis-network.min.js is dual-licensed under the Apache License 2.0 and MIT.