diff --git a/docs/embeddings/configuration/vectors.md b/docs/embeddings/configuration/vectors.md
index dce8c4e6..d4d4f88a 100644
--- a/docs/embeddings/configuration/vectors.md
+++ b/docs/embeddings/configuration/vectors.md
@@ -155,6 +155,18 @@ muvera:
 
 Settings to control the size of MUVERA fixed dimensional outputs. Default is 20 * 2^5 * 16 = 10,240 dimensions.
 
+### lemur
+```yaml
+lemur:
+    path: local artifact directory or Hugging Face Hub path
+```
+
+Loads a trained [LEMUR](../../../pipeline/train/lemur) fixed dimensional encoder. LEMUR artifacts are corpus-specific and must be
+created with `LemurTrainer` before configuring an embeddings index. Query vectors are summed learned features and document vectors
+are ordinary least squares weights over the artifact's stored token sample. Each artifact contains `config.json` and
+`model.safetensors`; the Safetensors file contains inference state only. See the trainer page for score filtering and Faiss exact/IVF
+search behavior.
+
 ### trust_remote_code
 ```yaml
 trust_remote_code: boolean
diff --git a/docs/pipeline/train/lemur.md b/docs/pipeline/train/lemur.md
new file mode 100644
index 00000000..20b7de0a
--- /dev/null
+++ b/docs/pipeline/train/lemur.md
@@ -0,0 +1,82 @@
+# LemurTrainer
+
+![pipeline](../../images/pipeline.png#only-light)
+![pipeline](../../images/pipeline-dark.png#only-dark)
+
+Trains a LEMUR fixed dimensional encoder for a late interaction model and corpus. Training is a separate pipeline step; the saved
+artifact is then loaded through the embeddings `vectors.lemur` configuration.
+
+## Example
+
+```python
+from txtai.pipeline import LemurTrainer
+
+corpus = [
+    "First document",
+    "Second document",
+    "Third document",
+]
+
+trainer = LemurTrainer()
+trainer(
+    "colbert-ir/colbertv2.0",
+    corpus,
+    "lemur-model",
+    epochs=100,
+    validation_split=0.1,
+)
+```
+
+`epochs` is required so the training mode is explicit. Use `epochs=100` for the quality-oriented MLP setting. This can take hours on
+a CPU. Use `epochs=0` to select deterministic random ELM features as a lower-cost fallback.
+
+The learn distribution defaults to `learn_category="query"`. Target documents are always encoded with the data encoder, so the
+default encodes selected corpus texts once as data and once as queries. Set `learn_category="data"` to reuse the data encodings, or
+pass a separate iterable of texts with `learn`.
+
+Set `corpus_subset_size` to a positive integer to sample that many raw corpus texts under `seed` before either encoding pass. The
+default is `None`, which uses every corpus text. `train_subset_size` and `learn_subset_size` apply later, after token vectors have
+already been created. `corpus_subset_size` applies to `data`; a separate `learn` iterable remains caller-sized.
+
+For trained MLP features, set `validation_split` to a fraction greater than zero and less than one to retain the epoch with the lowest
+held-out loss. The default is `0.0`, which preserves training-loss selection. The selected one-based epoch, loss and metric are
+available as `selected_epoch`, `selected_loss` and `selection_metric` on the fitted or reloaded encoder.
+
+The artifact contains `config.json` and `model.safetensors`. It stores the inference feature model, output-normalization statistics
+and token sample needed to encode documents added after training. The training-only output readout is not saved.
+
+Load the artifact in an embeddings configuration.
+
+```yaml
+embeddings:
+    path: colbert-ir/colbertv2.0
+    vectors:
+        lemur:
+            path: lemur-model
+```
+
+## Search behavior
+
+LEMUR approximates standardized MaxSim targets, so useful ranking scores can be negative. txtai's dense vector path L2-normalizes the
+fixed vectors and removes results with scores less than or equal to zero. This can change ordering and return fewer than the requested
+number of candidates compared with raw maximum inner-product search.
+
+The Faiss backend uses exact `IDMap,Flat` search through 5,000 rows and switches to an IVF index above that threshold. In the measured
+scifact run, default IVF reduced LEMUR NDCG@10 by 43% relative to exact search, compared with 25% for MUVERA. For LEMUR corpora above
+5,000 rows, either pin `faiss.components` to an exact index or tune the IVF settings for the corpus. The exact configuration is:
+
+```yaml
+embeddings:
+    path: colbert-ir/colbertv2.0
+    vectors:
+        lemur:
+            path: lemur-model
+    faiss:
+        components: IDMap,Flat
+```
+
+## Methods
+
+Python documentation for the pipeline.
+
+### ::: txtai.pipeline.LemurTrainer.__call__
diff --git a/mkdocs.yml b/mkdocs.yml
index 6af5c533..848322f6 100644
--- a/mkdocs.yml
+++ b/mkdocs.yml
@@ -109,6 +109,7 @@ nav:
             - Translation: pipeline/text/translation.md
         - Train:
             - HF ONNX: pipeline/train/hfonnx.md
+            - LEMUR Trainer: pipeline/train/lemur.md
             - ML ONNX: pipeline/train/mlonnx.md
             - Trainer: pipeline/train/trainer.md
     - Workflow: