gligabao (Python client)

Python client for the gligabao Go server.

from gligabao import Client

gb = Client("http://127.0.0.1:7878")
res = gb.remember("u1", "Alice works at Acme and lives in Paris.", kind="episode")  # kind defaults to "knowledge"
print(res["artifact_id"], res["edges"])

Each LLM stage of remember can run on a model of your choice with models (stages extract_graph, resolve_edges and, for images, extract_image); a stage you leave out uses the server's default model:

res = gb.remember("u1", "Alice works at Acme.", models={"extract_graph": "claude-sonnet-4-5"})  # resolve_edges: default

No LLM stage thinks by default ("none", so no thinking tokens are paid for); reasoning turns it on per stage ("low", "medium" or "high"). Each call is recorded by the server as an inference with its model and level, and each chunk summary references the run that wrote it:

res = gb.remember("u1", long_text, models={"summarize_chunk": "claude-haiku-5-5"}, reasoning={"summarize_chunk": "low"})
print(res["chunks"][0]["summary"])  # artifact_id, text, model, reasoning, inference_id

With web_search=True the server also searches the web for the latest world events about the key entities of the text and remembers what it finds (needs an Anthropic API key on the server; see docs/web-search.md of gligabao):

res = gb.remember("u1", "Apple and Microsoft are large companies.", web_search=True)
print([(w["entity"], w["query"], w["artifact_id"]) for w in res.get("web_searches", [])])

Timeseries can be stored with a text and anchored to what it yields (see docs/timeseries.md of gligabao). A series without anchors is anchored to the artifact; an entity or edge anchor is matched after the text is extracted, so it may name an entity the text will yield. attach_timeseries does the same for what is already stored:

from datetime import datetime, timezone

closes = [(datetime(2025, 1, 30, tzinfo=timezone.utc), 237.6), (datetime(2025, 1, 31, tzinfo=timezone.utc), 222.0)]
res = gb.remember("u1", "Apple fell after the earnings miss.", kind="episode", timeseries=[
    {"key": "AAPL.close", "unit": "USD", "points": closes,
     "anchors": [{"entity": {"name": "Apple", "type": "organization"}}]},
])
print([(s["key"], s["stats"]["trend"], s["anchors"]) for s in res["series"]])
gb.attach_timeseries("u1", [{"key": "AAPL.close", "points": closes, "anchors": [{"entity_id": 12}]}])

If the server was started with timeseries providers (a Go program embedding gligabao registers them), it can fetch series itself: an LLM reads the providers' descriptions and decides which provider is for which entity of the text.

res = gb.remember("u1", "Apple fell after the earnings miss.", kind="episode", fetch_timeseries={"providers": ["prices"], "before": "30d"})
print([(f["entity"], f["provider"], f["query"], [s["key"] for s in f.get("series", [])]) for f in res["series_fetches"]])

Add observe_timeseries=True to also derive a series_observation artifact from each stored series (a text generated from its stats, extracted into the graph like any text); they are listed under series_observations.

An image is remembered by the image pipeline: the server's LLM reads the text in it (title and text blocks) and classifies the whole image by a theme (plot, graph, chart, table, document, ...). Pass the raw bytes, or the path of a JPEG, PNG, GIF or WebP file (up to 5 MiB); text and kind do not apply:

res = gb.remember("u1", image="revenue.png", dataset="charts")
print(res["theme"], res["title"], [b["text"] for b in res["text_blocks"]])

Install for development: pip install -e ".[dev]", then pytest.