Inspection: Embedding (non-SAE)

Geometry and activation tools for encoder models: layer drift, isotropy, OOD separation, attention, and token attribution without sparse autoencoder decomposition. Requires embedding mode: load an embedding model first.

Prerequisiteaquin load --model gte-small (or all-minilm-l6, bge-small-en)

10 commands

aquin embed-layer-drift

agent tool: run_embed_layer_drift

Traces the L2 norm of the pooled embedding at each encoder layer. The steepest change marks where representation shift concentrates. Optional ref_text overlays a second curve for comparison.

Flag	Description
--text*	Primary input text.
--ref_text	Reference text for overlay comparison.

example

aquin embed-layer-analysis

agent tool: run_embed_layer_analysis

Combined pass: layer drift, isotropy, OOD separation, and embedding consistency in one call. Prefer this when you want a full layer health report without running four separate commands.

Flag	Description
--text*	Primary text.
--ref_text	Reference for drift overlay.
--texts	JSON array for isotropy / space analysis.
--in_texts / --ood_texts	In-distribution and OOD sets.
--paraphrases	Paraphrases for consistency scoring.

example

aquin embed-isotropy

agent tool: run_embed_isotropy

Computes isotropy (how uniformly embeddings occupy directions in space) and spectral entropy per layer. Low isotropy means the space is anisotropic: a few directions dominate, which hurts retrieval diversity.

Flag	Description
--texts*	JSON array of strings to encode.

example

aquin embed-ood

agent tool: run_embed_ood

Compares in-distribution vs out-of-distribution text embeddings at each layer. Reports separation score: how well the model distinguishes familiar from unfamiliar input geometry.

Flag	Description
--in_texts*	JSON array of in-domain texts.
--ood_texts*	JSON array of OOD texts.

example

aquin embed-attention

agent tool: run_embed_attention

Extracts per-head attention matrices across all encoder layers for a single input. Shows which tokens the encoder attends to when building the pooled embedding.

Flag	Description
--text*	Input text.

example

aquin embed-matrix

agent tool: run_embed_matrix

Encodes multiple texts and renders their N×N cosine similarity matrix. Useful for sanity-checking whether related sentences cluster and unrelated ones separate.

Flag	Description
--texts*	JSON array of strings.

example

aquin embed-space

agent tool: run_embed_space_analysis

Measures anisotropy (variance concentration along principal axes) and intrinsic dimensionality of the embedding cloud. High anisotropy indicates collapse onto a few directions.

Flag	Description
--texts*	JSON array of strings.

example

aquin embed-attribution

agent tool: run_embed_attribution

Integrated gradients token attribution on the final embedding. Scores each input token by how much it contributes to the embedding vector, the embedding analogue of LLM token attribution.

Flag	Description
--text*	Input text.

example

aquin embed-perturbation

agent tool: run_embed_perturbation

Perturbs individual tokens (mask, swap, delete) and measures cosine shift in the output embedding. Identifies tokens the representation is most sensitive to.

Flag	Description
--text*	Input text.

example

aquin embed-retrieval

agent tool: run_embed_retrieval

Encodes a query and corpus, ranks documents by cosine similarity, and reports Recall@k metrics. Tests whether the loaded encoder retrieves the right documents for a query set.

Flag	Description
--query	Single query string.
--corpus	JSON array of document strings.

example