Inspection: Embedding (non-SAE)

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

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

7 commands

aquin check layer

agent tool: run_embed_layer_analysis

Layer health in one pass: pooled-embedding drift across encoder layers (where representation shift concentrates), plus isotropy, OOD separation, and paraphrase consistency. Same command name as LLM mode; with an embedding model loaded this is the embed path.

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.
--check	Save layer-analysis-check.json and layer-analysis-check.png in the current directory.

example

aquin check attention

agent tool: run_embed_attention

Extracts per-head attention matrices across all encoder layers for a single input. Same command as LLM mode; with an embedding model loaded this is the embed path. --text is required.

Flag	Description
--text*	Input text.
--check	Save attention-check.json and attention-check.png in the current directory.

example

aquin 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.
--check	Save matrix-check.json and matrix-check.png in the current directory.

example

aquin 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.
--check	Save space-check.json and space-check.png in the current directory.

example

aquin check 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.
--check	Save attribution-check.json and attribution-check.png in the current directory.

example

aquin check 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.
--check	Save perturbation-check.json and perturbation-check.png in the current directory.

example

aquin retrieval

agent tool: run_embed_retrieval

Encodes one query and a candidate corpus, then ranks every passage by cosine similarity. Returns each passage, score, and rank; it does not compute Recall@k/MRR without relevance labels.

Flag	Description
--query	Single query string.
--corpus	JSON array of document strings.
--check	Save retrieval-check.json and retrieval-check.png in the current directory.

example