v0.3.1

Added

• bee.requeue(p_table, p_column, p_row_pk) queues one existing row from its current sources, even when a model result already exists, without changing human overrides. bee.retry_dead_row(p_table, p_column, p_row_pk) retries only that row's dead job instead of every dead job in the column. Both return whether a job was queued and require an explicit EXECUTE grant for separate application roles.

Upgrading

• Extension version 0.17 (0017_row_queue_management.sql): run pgbee install, or ALTER EXTENSION pgbee UPDATE. Grant application roles USAGE ON SCHEMA bee and EXECUTE on the new functions as needed. The worker contract does not change.

v0.3.0

Added

• bee.definition(p_table, p_column): the current definition of a derived column (version, backend, model, prompt, output schema, backend config, config) without the queue counters of bee.columns, cheap enough to read on every request.

Changed

• Results of vector and halfvec columns no longer copy the vector into bee.result.value, which stays null: the column holds the value, and the copy took ten times its space as jsonb.

Fixed

• A deleted row kept a current result forever, so its lineage was never pruned, and a row deleted and inserted again with the same key and sources was never computed. Deleting a row, or truncating the table, now retires its current result to the superseded lineage and drops its live jobs.

Upgrading

• Extension version 0.16 (0016_forget_deleted_rows.sql): run pgbee install, or ALTER EXTENSION pgbee UPDATE. The update installs two triggers on every table with a derived column (bee_forget_<column>, bee_forget_all_<column>), retires the current results of rows that no longer exist with one anti-join per table, and clears the vectors stored in the lineage: on large tables and lineages it takes a while and locks the rows it updates. The worker contract does not change.

v0.2.0

Added

• Output type halfvec for the embedding backend: bee.add_column(..., 'halfvec', 'embedding', p_output_schema => '{"dimensions": N}') writes half-precision vectors to a pgvector halfvec(N) column. Half the storage of vector, and HNSW indexes it up to 4000 dimensions instead of 2000, enough for 3072-dimension models. Needs pgvector 0.7 or later; an existing halfvec(N) column is used as the target.

Upgrading

• Extension version 0.15 (0015_halfvec.sql): run pgbee install, or ALTER EXTENSION pgbee UPDATE where pgbee is installed as an extension. The worker contract does not change; a worker from 0.1.0 already writes halfvec columns.

v0.1.0

First public release.

Added

• Derived columns declared in SQL: bee.add_column sets a prompt, a model, the source columns and an output type (enum, text, boolean, integer, numeric, jsonb, vector). Inserts and updates of the sources enqueue a job in the same transaction; the value is written back by a worker.
• Backends: llm (structured output with self-reported confidence), decision (typed questions to TypeSafe Jev with class probabilities; the decision columns of a row share one call, up to 16 questions), embedding (batched vectors, needs pgvector) and custom (jobs claimed by your own worker).
• Versions and lineage: bee.update_column creates a new version and recomputes only the rows computed with an older one. Every value is kept in bee.result with its model, prompt version, confidence, tokens, cost and latency; a value computed while the row or the version changed is recorded under the version it was computed with and the row is queued again.
• Human overrides: a value written by hand is never overwritten (override_policy pin, or until_source_change); bee.unpin or setting the cell to NULL gives it back to the model. Writing the value a cell already holds is not an override.
• Review queue: values below confidence_threshold are listed in bee.needs_review; low_confidence_policy hold keeps them out of the column.
• Budgets per column (budget_usd per day, month or in total), enforced when jobs are claimed, with bee.budgets, bee.spent and bee.cost_by_column. Token prices per column (input_usd_per_mtok, output_usd_per_mtok) for providers that report no cost.
• Retries with exponential backoff, dead jobs in bee.dead_jobs and bee.retry_dead, abandoned jobs reclaimed after a timeout.
• Backfill in chunks: declaring a column on a table of a million rows takes a fraction of a second and does not block writers. bee.disable and bee.enable for bulk loads.
• Lineage retention per column (lineage_retention_days), pruned in batches by bee.prune.
• Least-privilege worker: the bee_worker role can only call the contract functions and read the views; it has no access to user tables.
• Worker contract: bee.claim_jobs, bee.complete_job, bee.fail_job, bee.reclaim_stale, bee.prune, and the bee_jobs NOTIFY channel. Any process that speaks it is a valid worker.
• Two ways to install, generated from the same SQL files: pgbee install on any PostgreSQL 15 to 18, managed services included, or CREATE EXTENSION pgbee from pgbee extension-files on self-hosted servers, with ALTER EXTENSION pgbee UPDATE and pg_dump support.
• Reference worker pgbee (Python 3.13): pgbee run with one lane per backend so fast models never wait for slow ones, --backends to split backends across machines, --drain and --max-seconds for cron and serverless schedulers, pgbee status, pgbee install, pgbee extension-files.
• Model providers: OpenRouter (all backends, reported cost) or any OpenAI-compatible endpoint such as OpenAI, Azure OpenAI, Ollama or vLLM (llm and embedding, standard parameters only), chosen with OPENROUTER_API_KEY, OPENAI_API_KEY, OPENAI_BASE_URL or PGBEE_PROVIDER.
• Docker images: ghcr.io/essedev/pgbee-worker and ghcr.io/essedev/pgbee-postgres (PostgreSQL 17 and 18 with the extension available), and a docker compose quickstart.
• Benchmarks anyone can rerun without cost in bench/: a failure test against two app-side designs, scale on a million rows, lanes against a single loop.

Upgrading

• Nothing to upgrade from. This release ships extension version 0.14 (SQL files 0001 to 0014). Install with pgbee install, or with CREATE EXTENSION pgbee after copying the files from pgbee extension-files (also attached to the GitHub release).
• On managed services (tested on Neon) connect the worker directly, not through a transaction pooler: the pooler drops the notifications that wake the worker, which then only polls every PGBEE_POLL_INTERVAL_SECONDS.