qombra

Python client for Qombra — data analysis, AI-guided preprocessing, model training on QBrain, inference, zero-shot time series forecasting, and SHAP explainability, all running on the Qombra platform through your account.

QBrain is Qombra's proprietary tabular foundation model: it reads the structure of your data directly, so fit() needs a dataframe and a target column — no architecture to pick, no hyperparameters to tune.

pip install qombra

Quickstart

import pandas as pd
import qombra

qombra.login()   # opens the browser; approve the SDK session (valid 12 hours)

df = pd.read_csv("customers.csv")

# 1. Dataset statistics (analyzes a ≤1000-row sample with the fewest NaNs)
report = qombra.analyze(df)
print(report.summary, report.warnings)

# 2. Preprocessing with a natural-language instruction
result = qombra.preprocessing(df, "drop duplicate rows and outliers in price")
print(result)          # summary, actions, warnings
clean_df = result.df

# 3. Training — the model stays on the server, addressed by id
model = qombra.fit(clean_df, target="churn_30d")
print(model.id, model.metrics)
# Optional: pick the evaluation metric and the compute budget yourself
model = qombra.fit(clean_df, target="churn_30d", metric="roc_auc", effort="high")
# Optional: fine-tune QBrain on your data (slower; kept only if it improves the model)
model = qombra.fit(clean_df, target="churn_30d", finetuning=True)

# 4. Inference
predictions = model.predict(clean_df.head(100))

# 5. Explainability (SHAP)
print(model.explain())

qombra.logout()  # revoke the session token

Later, in another session — no retraining

import qombra

qombra.login()
model = qombra.Model.from_id("«the model id from earlier»")
predictions = model.predict(new_rows)

Session management

There are two ways to authenticate, for the two ways people use the API.

Interactive — qombra.login(). Opens a browser window where you sign in on the web app and approve the SDK; the resulting session lives in your OS keyring and expires after 12 hours. Close it explicitly, or scope it with with (leaving the block ends the session):

qombra.login()
with qombra.Qombra() as client:
    model = client.fit(df, target="price")
    print(client.whoami())   # remaining quotas
# session ended here

No browser available (SSH, CI)? Use qombra.login(headless=True) and copy-paste the code shown on the consent page.

Production — an API key. For systems that must run unattended, a key never expires and needs no browser. Create one in the Qombra app under Account → API keys, then put it in the environment or in a .env file in your project (the environment wins when both are set):

export QOMBRA_API_KEY=qbk_…

import qombra
preds = qombra.predict(model_id, new_rows)   # no login() call

The key is shown once at creation — store it in your secret manager. It stays valid until you revoke it in the same place; rotate by creating a new key, deploying it, then revoking the old one. close() never revokes an API key, so a with block cannot take your service offline.

API keys are available on accounts enabled for production access; the tab appears in the app once Qombra switches it on for you.

The full pipeline in one call

result = qombra.auto_run(df, "Predict which customers churn in the next 30 days")
print(result)                     # phases, target, test metric
preds = result.model.predict(new_rows)

auto_run drives the same agent workflow as the web app (ingest → preprocessing → target confirmation → training) without a human in the loop. Expect minutes to hours; the created analysis is fully browsable in the web app afterwards.

Forecasting time series

QBrain forecasts zero-shot: hand it the history and it returns the forecast — no training step, no model id, nothing to delete afterwards.

# sales: one row per store and day — store, date, units, price, promo
fc = qombra.forecast(
    sales, target="units", horizon=14,
    timestamp_column="date", id_column="store",
    future=planned,                       # store, date, price, promo for the next 14 days (optional)
    groups={"north": ["s1", "s2", "s3"]}, # related series inform each other (optional)
    non_negative=True,                    # units can't go below zero
)
fc.frame.head()          # store, date, target, forecast, forecast_q05, forecast_q50, forecast_q95
fc.wide()                # one column per store, one row per forecast day
fc.wide(quantile=0.95)   # the upper band
print(fc.warnings)       # duplicates dropped, gaps filled, series skipped, ...

• Long format. One row per observation; every column that is not the id, the timestamp or a target is a past covariate (numeric or categorical). Leave id_column unset for a single series.
• Future covariates go in future: the id and timestamp columns plus the covariate columns, one row per series per forecast step, starting right after each series' last observation. Columns of df absent from future are treated as past-only.
• Frequency is inferred per series; pass freq="W-MON" (any pandas frequency alias) when the data is too gappy to infer from. Gaps are filled, duplicate timestamps collapsed (last wins), timezone-aware timestamps moved to UTC.
• Quantiles: any number of levels between 0.01 and 0.99; 0.5 is always included and is the forecast column. Each level adds a column, so many levels count against the forecast-values limit.
• Several targets (target=["units", "revenue"]) are forecast jointly per series and share that series' covariates. A target that needs covariates of its own is better modelled as its own series (its own id), at the price of being forecast independently.
• Limits per call: 10,000 series · 100 value columns per series · horizon 1,024 steps · 20,000,000 data points (history rows × value columns, after each series is truncated to its most recent 8,192 steps) · 10,000,000 forecast values (series × horizon × targets × (1 + quantile levels)).

A forecast that outlives your client-side timeout keeps running server-side; collect it with qombra.forecast_result(e.job_id) from the JobTimeoutError. Ctrl+C or qombra.stop(job_id) cancels it.

Managing stored artifacts

qombra.list_models()                     # all trained models in your account
qombra.delete_model(model)               # irreversible
qombra.list_preprocessing_results()
qombra.delete_preprocessing_result(job_id)

Async usage

In asyncio applications (web services, agents), use AsyncQombra — it mirrors the whole client surface as coroutines, so uploads and job polling yield to the event loop instead of blocking it:

import qombra

async with qombra.AsyncQombra() as client:       # or AsyncQombra(api_key=...)
    model = await client.fit(train_df, target="churn")
    predictions = await model.predict(new_df)    # an AsyncModel — awaitable too
    fc = await client.forecast(sales, target="units", horizon=14, id_column="store")

Everything else is identical to the sync client: endpoints, credentials, validation, retries, timeouts, warnings, result types and quotas. Two differences in shape: instant predictions are a method on AsyncModel (await model.instant_predictions()) rather than a property, and the browser sign-in stays synchronous — run qombra.login() before entering the event loop, or pass api_key=. Cancelling the task awaiting fit() / forecast() / auto_run() sends the same best-effort server-side cancel as Ctrl+C does in the sync client.

Error handling

All errors derive from qombra.QombraError:

try:
    model = qombra.fit(df, target="revenue")
except qombra.AuthenticationError:
    qombra.login()                       # token expired (12h) — sign in again
except qombra.QuotaExceededError as e:
    print("Usage limit reached:", e)
except qombra.ValidationError as e:
    print("Bad input:", e.code, e)
except qombra.JobTimeoutError as e:
    print("Still training server-side, job:", e.job_id)

Notable classes: AuthenticationError, QuotaExceededError, ValidationError (with a machine-readable .code), PayloadTooLargeError, NotFoundError, JobFailedError, JobTimeoutError, NetworkError, ServerError.

Data format & metering

Dataframes travel as parquet with plain scalar columns (numbers, booleans, strings, dates, timestamps, decimals; pandas categoricals are fine) — cast mixed-type object columns before upload. Uploads are size-capped and calls are rate-limited: an oversized upload raises PayloadTooLargeError (reduce or batch it), rapid-fire calls raise RateLimitedError.

All SDK usage counts toward the same account limits the web app uses:

• preprocessing spends LLM output tokens (the AI agent's work) plus one chat message, charged once per call.
• fit creates one analysis, counted against your analysis limit. QBrain is not an LLM, so training spends no output tokens.
• auto_run spends all three: one analysis, plus the chat messages and output tokens the agent consumes across the run.
• analyze, predict, forecast, explain and the listing calls spend no quota; they are rate-limited instead (predict and forecast share one hourly limit).

Check what is left with qombra.whoami().

Support

Questions and issues: www.qombra.com — or reach out at hey@qombra.com.