Triplet Embeddings (removed) - Cognee Documentation

Triplet embeddings were removed from Cognee. Cognify no longer writes the Triplet_text vector collection, and the create_triplet_embeddings pipeline, the get_triplet_datapoints memify task, and TripletRetriever no longer exist.

What to use instead

Search: use SearchType.GRAPH_COMPLETION (or the default HYBRID_COMPLETION). Graph completion searches edges through the relationship-name index that cognify already writes, so no extra indexing step is needed.
Enrichment: use improve() to enrich a dataset and bridge session memory, or run_custom_pipeline() to run your own tasks over it.

What still works for one release

These surfaces stay as deprecated shims that raise a FutureWarning, and are removed in the next release:

Surface	Behavior now
`SearchType.TRIPLET_COMPLETION`	Runs as `GRAPH_COMPLETION` in `search()` and `recall()`.
`cognee.memify()` / `POST /api/v1/memify`	Named or given tasks still run; a call without tasks runs nothing and returns `{}`. The task name `get_triplet_datapoints` is rejected with `CogneeValidationError` (422 over REST). See memify().
`cognee-cli memify`	Prints a deprecation notice and exits. Use `cognee-cli improve`.
`improve(extraction_tasks=..., enrichment_tasks=...)`	The tasks still run, as an extra `custom_tasks` stage. See improve().
`add_data_points(embed_triplets=...)`	Accepted and ignored.

The TRIPLET_EMBEDDING environment variable is ignored; when it is set, Cognee logs a warning once per process. Existing Triplet_text collections are not dropped automatically. Deleting data still removes its rows from a legacy Triplet_text collection if one exists.

Improve

Understand the current improvement workflow

Recall

Query the graph with v1.0 retrieval

⌘I