Configuration

The gligabao.toml config file, API keys, vector store, prices and eval/admin modes.

Configuration is read from build/gligabao.toml (override with --config).

Settings

Key	Default	Purpose
port	7878	HTTP port
anthropic_api_key	ANTHROPIC_API_KEY	LLM for remember/query
openai_api_key	OPENAI_API_KEY	Embeddings (without it no vectors are stored)
pinecone_api_key	PINECONE_API_KEY	Hosted vector store
db_backend	"sqlite"	"sqlite" or "postgres" (with postgres_dsn), see Database
vector_store	"chromem"	"chromem" (embedded) or "pinecone" (hosted)
embedding_dimensions	1536	Vector size
pinecone_index_host	PINECONE_INDEX_HOST	Data-plane host of the Pinecone index
chunk_size	1024	Tokens per chunk, see Chunking
chunk_overlap_percent	20	Overlap between chunks
jev_api_key	JEV_API_KEY	Enables Jev, TypeSafe’s system 1 model (jev_base_url and jev_model are optional overrides)
sidecar_url	SIDECAR_URL, else http://127.0.0.1:8787	Base URL of the Python sidecar’s REST API, the only service the Go server calls for training jobs; the sidecar proxies to Modal
eval	false	Register the eval endpoints
admin_ui	false	Serve the admin panel
geo	false	Make remember find the places among the entities it extracts and store their coordinates (gb_entity_geo), using the public Nominatim geocoder of OpenStreetMap; costs one more LLM call per text and sends the names of places to the geocoder. serve --geo turns it on. See Geo
mcp	false	Serve remember and query as MCP tools at /v1/gb/mcp (streamable HTTP), so MCP clients such as Claude Desktop or Claude Code can use gligabao as a memory. No authentication: enable it only on a trusted network. serve --mcp turns it on. See MCP

The chromem store fixes the embedding size when its directory is created, and a Pinecone index must match it. Details: Vector storage.

Prices

A [prices."<model-prefix>"] table (input, output, optional cache_write/cache_read, USD per 1M tokens) overrides or extends the built-in model prices used to compute the cost of each LLM call; the longest matching prefix wins.

[prices."claude-haiku"]
input = 1.0
output = 5.0

Every model invocation (LLM call, embedding request, web search, Jev classification) is a row of the gb_inferences table, with the model, the thinking level, the tokens and the cost. See Inferences.

Jev

When a Jev key is set, remember first classifies the text as simple/medium/complex in the number of entities it will yield and logs it with Jev’s confidence and the call’s latency. See Remember.

Eval mode

With serve --eval (or eval = true) the server also registers POST /v1/gb/eval/remember and POST /v1/gb/eval/query. They take the same bodies as the regular endpoints, record the call as an ATIF trajectory, store it as pretty-printed JSON in an atif/ directory next to the SQLite file, and return the full trajectory (and its trajectory_id) in the response. They are not registered otherwise, and the regular endpoints never generate trajectories. go run ./evals/knowledge runs through these endpoints (-atif <dir> also keeps the trajectories as files). See Evals.

Admin UI

admin_ui/ holds a small admin panel (TypeScript, SCSS, HTML). ./run_build.sh builds it (needs node/npm) and embeds admin_ui/dist into the binary with go:embed; pass --no-admin-ui (or BUILD_ADMIN_UI=false) to skip it. With admin_ui = true (or serve --admin-ui; default false) the server serves it at /v1/gb/admin/. If the assets were not embedded, the server logs a warning and does not register the handlers.

The panel has views for the databases (main), evals, timeseries, inferences, remember runs and training: the training view lists the training datasets (the manifests in the training datasets directory), uploads datasets and runs Gemma-4 fine-tuning jobs on the sidecar, with a dataset picker, the list of jobs, a live event feed and a loss graph (see Training). It needs sidecar_url to reach the sidecar, which proxies to Modal. A sidecar: up / sidecar: down badge in the top bar shows whether the sidecar answers its health check. The training view also lists the training backends (Modal: healthy, app deployed or not) and can deploy, cancel, stop and inspect the volumes of the Modal app, with a live build log: see Modal app lifecycle.

A program that embeds gligabao can extend the admin UI with its own views (a button in the nav on the left) and with buttons, modals and blocks inside the existing views, with plain JavaScript and no build step: see Admin UI extensions.