---

Agents build RAG pipelines, semantic search and recommendation retrieval every day, but they cannot see the embedding space they are searching.

When retrieval goes wrong it rarely looks like a bug. It looks like:

• "the same three chunks come back for everything"
• "every score is 0.8-something"
• "top-k is full of copies"
• "brute force finds it and HNSW does not"

Each of those failures has a known, deterministic diagnostic in the literature. vecscope turns them into one command and one JSON:

• measurements
• ok / warn / fail verdicts
• an explanation an agent can act on

It runs locally on NumPy, needs no model or network, and reports the same numbers every time.

vecscope diagnoses the vector search implementation: the geometry the model produced, the hygiene of the corpus, the distance metric, the index and the query pipeline.

It does not measure relevance. Whether the right documents come back for a question needs labeled queries, and a clean audit is evidence that search is not sabotaging retrieval, not proof that retrieval is good.

Quick start

1. Install

pip install vecscope

2. Get a file of embeddings

vecscope reads a file with one embedding vector per row: the (n, dim) array a model's encode returns, saved as .npy, .parquet, .jsonl, .json or .csv (all formats).

import numpy as np
embeddings = model.encode(texts)          # shape (n, dim), e.g. (3000, 128)
np.save("embeddings.npy", embeddings)

No embeddings at hand? Generate three synthetic matrices (healthy, anisotropic, collapsed):

python -c "import urllib.request as u; u.urlretrieve('https://raw.githubusercontent.com/076923/vecscope/main/examples/make_examples.py', 'make_examples.py')"
python make_examples.py

3. Run the audit

vecscope audit embeddings.npy

vecscope x.y.z  n=3000 dim=128 k=10  verdict=FAIL
  [ok  ] hubness                skewness=1.1661  (ok < 2, warn >= 2, fail >= 4)
                               k-occurrence skewness 1.17; top 1% of items take 3.9% of all neighbor slots (max N_k=51, expected 10).
  [WARN] antihubs               antihub_excess=0.0523  (ok < 0.05, warn >= 0.05, fail >= 0.2)
                               5.2% of items are never anyone's neighbor (random data at k=10 would give 0.0%); the excess is unreachable through k-NN retrieval.
                               try: inspect a sample of never-retrieved items; drop empty inputs, then apply the hubness fix
  [FAIL] anisotropy             mean_random_cosine=0.8732  (ok < 0.5, warn >= 0.5, fail >= 0.85)
                               random pairs have mean cosine 0.873 (isotropic ~0); mean unit vector has norm 0.93; centering would bring the cosine to -0.000.
                               try: subtract the corpus mean vector and re-normalize, applying the same transform to queries; details.mean_random_cosine_after_centering shows the expected result
  [ok  ] effective_rank         effective_rank_ratio=0.5407  (ok >= 0.1, warn < 0.1, fail < 0.05)
                               effective rank 69.2 of 128 attainable (128 dims; 54.1%); low values mean most dimensions carry nothing.
  [ok  ] collapse               near_duplicate_ratio=0.0000  (ok < 0.02, warn >= 0.02, fail >= 0.1)
                               0.0% of items have a neighbor with cosine >= 0.99 or an exact copy (exact duplicates 0.0%); no row has a near-duplicate.
  [ok  ] rogue_dimensions       top1_dim_variance_ratio=1.3070  (ok < 10, warn >= 10, fail >= 50)
                               dimension 94 alone holds 1.0% of variance, 1.3x the uniform share of 0.78%; top 3 dims hold 3.0%. Rogue dimensions dominate cosine.
  [ok  ] norms                  norm_cv=0.0000  (ok < 0.1, warn >= 0.1, fail >= 0.3)
                               vector norms vary with CV 0.000 (median 1.000, range 1.000-1.000); 0.00% are zero vectors. Unequal norms bias inner-product search toward long vectors.
  [ok  ] metric_mismatch        topk_overlap=1.0000  (ok >= 0.9, warn < 0.9, fail < 0.7)
                               cosine and inner-product top-10 share 100.0% of neighbors; 100.0% of items get identical lists. Low overlap means the index metric changes the results.
  [ok  ] distance_concentration contrast_over_noise=2.2784  (ok >= 1.4, warn < 1.4, fail < 1.1)
                               on 3000 rows a random item is 3.27x farther than the nearest neighbor (random distance 0.127); isotropic noise at 128-d and the same rows would give 1.44x, so the contrast is 2.28x that of noise. On all rows the contrast is 3.27x (median NN distance 0.039). Near 1.0 means distances carry no signal.
  [ok  ] outliers               outlier_ratio=0.0030  (ok < 0.01, warn >= 0.01, fail >= 0.05)
                               0.30% of items have no neighbor within cosine distance 0.050 (median 0.039); see details.sample_ids. Typically junk, empty or foreign-language content.
  [ok  ] connectivity           island_ratio=0.0000  (ok < 0.01, warn >= 0.01, fail >= 0.1)
                               0.00% of items sit in 0 small island(s) of the k-NN graph (fewer than 15 items; 60 components in total). Graph-based ANN (HNSW) attaches islands to the wrong region or never reaches them.

4. Read the result

• The first line gives the overall verdict: the worst verdict among the findings.
• Each finding shows the measured value, the range that would be ok, and a sentence saying what the number means for retrieval. Findings that are not ok add a try: line with the smallest change to test first.
• The exit code is 1 on FAIL, so the same command works as a CI gate.
• Add --json to get the same evidence as JSON for an agent or a script.

What vecscope measures

Finding	Metric	ok	warn	fail	What it means when it fails
hubness	skewness of the k-occurrence distribution	< 2.0	2.0 to 4.0	≥ 4.0	a few items appear in everyone's k-NN list and will top every query
antihubs	share of items never retrieved as a neighbor, above the random baseline e^-k	< 5%	5 to 20%	≥ 20%	part of the corpus is unreachable through k-NN
anisotropy	mean cosine of random pairs	< 0.85	0.85 to 0.95	≥ 0.95	vectors crowd into a narrow cone, so scores compress into a band; good retrievers sit well below this without losing nDCG, so only extreme values are flagged
effective_rank	participation ratio of the covariance spectrum / attainable rank (dim, or rows − 1 when rows do not exceed dims)	≥ 0.10	0.05 to 0.10	< 0.05	most dimensions carry nothing
collapse	share of rows with a neighbor at cosine ≥ 0.99 or an exact copy (zero vectors included); details give the duplicate groups and sample pairs	< 2%	2 to 10%	≥ 10%	top-k is spent on copies of the same item
ann_recall	recall@k of ANN neighbors vs exact k-NN, counted by distance so that exact ties (duplicates) are interchangeable; the tie slack is 1e-6 + 2e-6 x distance, two to three times float32 rounding	≥ 0.95	0.85 to 0.95	< 0.85	the index, not the model, is losing results
rogue_dimensions	variance of the largest dimension as a multiple of the uniform 1/dim share	< 10x	10 to 50x	≥ 50x	a few coordinates decide every cosine score
norms	coefficient of variation of vector norms	< 0.10	0.10 to 0.30	≥ 0.30	inner-product search favors long vectors; zero vectors are reported too
metric_mismatch	top-k overlap between cosine and inner product	≥ 0.90	0.70 to 0.90	< 0.70	the index metric changes which results come back
distance_concentration	relative contrast (random distance / nearest-neighbor distance) divided by that of isotropic noise at the same dimension and row count	≥ 1.4	1.1 to 1.4	< 1.1	the nearest neighbor is barely closer than a random item; 1.0 is pure noise
outliers	share of items with an abnormally far nearest neighbor	< 1%	1 to 5%	≥ 5%	junk, empty or foreign items; ids are listed in details
connectivity	share of items in small k-NN graph islands	< 1%	1 to 10%	≥ 10%	self-contained groups too small to be clusters; HNSW attaches them badly or never reaches them
query_gap	median query-to-corpus nearest distance divided by the median corpus-to-corpus nearest distance	< 2.6	2.6 to 3.0	≥ 3.0	queries and corpus do not share a space, typically queries embedded by a different model (5 to 10). Pipelines that reproduce published nDCG@10 sit at 0.94 to 2.52; a missing query prefix does not reliably move it
query_shift	share of queries farther from the corpus than the median + 3 MAD of a reference query set (needs a baseline)	< 20%	20 to 30%	≥ 30%	the query pipeline moved since the reference: another model or version, or traffic on topics the corpus does not cover. Healthy query sets score 0 to 9%
ood_queries	deprecated, read query_gap: share of queries farther from the corpus than 99% of corpus items are from each other	< 50%	50 to 80%	≥ 80%	same question as query_gap, but it counts the normal distance between short queries and long passages too (good models score 0 to 65% on BEIR) and false-alarmed on 4 of 18 reproduced pipelines
query_hubness	skewness of corpus k-occurrence over query results	< 4.0	4.0 to 10.0	≥ 10.0	the same few items answer most queries; the value tracks nDCG@10 across models (Spearman −0.5 to −0.9), wrong-language models score 11 to 30. It grows with corpus size: good English models score 1.5 to 3.4 up to 26k documents and 4.0 to 4.8 on 58k, so on large corpora compare against a baseline

Every finding in the JSON carries its healthy_range, warn_at and fail_at, plus a details block with the raw numbers. For anisotropy that block includes the cosine the same vectors would have after centering, so the agent can see whether centering alone fixes the problem.

Some findings appear only when you pass their inputs:

• ann_recall: an ANN neighbor file with --ann
• query_gap, query_hubness and ood_queries: query embeddings with --queries
• query_shift: query embeddings, and --baseline also pointing at an earlier audit run with queries

Each finding also carries likely_causes, suggested_action and verify_by. They are hypotheses, not prescriptions:

• apply the smallest change
• re-audit
• compare retrieval on real queries before adopting it

The agent skill spells out that loop.

Thresholds

The thresholds are defaults:

• calibrated on seven public models
• re-checked against 19 models on nine labeled datasets in two languages (verification)

They are a starting point for diagnosis, not a universal quality bar. Embedding spaces differ by model, language, domain and dimensionality, so a warn on a good model is normal and a clean audit does not guarantee good search.

Override any pair per run, or derive them from your own data with vecscope calibrate:

vecscope audit vectors.npy --thresholds my_thresholds.json
# my_thresholds.json: {"anisotropy.mean_random_cosine": [0.4, 0.7], "hubness.skewness": [2.5, 5]}

The keys are listed in src/vecscope/evidence.py (THRESHOLDS) and echoed in every JSON result under thresholds.

Queries and clusters

Pass the query embeddings too and vecscope checks whether queries live where the corpus lives.

• A different model on the query side shows up as query_gap.
• The same few chunks answering everything shows up as query_hubness, which is also what an embedding model that does not cover the corpus language raises.
• With a baseline from a run whose queries were known to be healthy, query_shift reports how many of today's queries moved away from the corpus compared with then. This also catches a partial change (some traffic on another model or on new topics).

vecscope audit corpus.npy --queries queries.npy
vecscope audit corpus.npy --queries good_queries.npy --json > baseline.json
vecscope audit corpus.npy --queries todays_queries.npy --baseline baseline.json   # adds query_shift

A missing query prefix (bge, e5) is not something these findings detect: on SciFact and NFCorpus it changed nDCG@10 by -0.014 to +0.007 and moved the query distances in either direction (verification). Check the prefix in the code that embeds queries.

Add --clusters 10 to append a per-cluster map:

• size
• tightness
• near-duplicate share
• how much of the hub traffic each cluster takes
• sample ids the agent can look up

It carries no verdicts; it tells the agent where a problem lives.

  clusters (3):
     id   size  share  cos2ctr  neardup  hubshare  sample_ids
      2   1113  37.1%    0.217     0.0%     37.0%  [1282, 1611, 496, 283, 2041]
      0   1023  34.1%    0.232     0.0%     34.2%  [1855, 2932, 1786, 1025, 2476]
      1    864  28.8%    0.222     0.0%     28.9%  [2505, 647, 2740, 2499, 2965]

Index sweep

vecscope sweep builds HNSW and IVF indexes with a small parameter grid and reports, for each setting on this corpus:

• recall@k against exact k-NN
• build time
• queries per second
• single-query latency percentiles (p50/p95/p99)
• the serialized index size

It then recommends the fastest setting that reaches the target recall (default 0.95) and reports its margin over the target.

A margin under 0.02 marks the recommendation tight and names the fastest setting with real headroom, because recall at fixed parameters falls as the corpus grows: a setting that just clears the target on a 50k-row sample usually misses it on the full corpus.

It needs the faiss-cpu extra.

pip install "vecscope[index]"
vecscope sweep embeddings.npy -k 10 --target-recall 0.95

  index params                                        recall  build s       qps  p50 ms  p95 ms     MiB  dev
 *hnsw  M=8, ef_construction=100, ef_search=16         0.987     0.35    237954   0.034   0.051     0.6  cpu
 *hnsw  M=16, ef_construction=100, ef_search=32        0.998     1.23    165139   0.084   0.108     0.8  cpu
 *ivf   nlist=76, nprobe=64                            0.999     0.14     33251   0.087   0.095     0.4  cpu
  recommendation: hnsw (M=8, ef_construction=100, ef_search=16)  recall=0.987  qps=237954  p95=0.051ms  index=0.6MiB  margin=+0.037  (fastest configuration with recall@5 >= 0.95)
  note: recall is measured on 200 queries drawn from the corpus itself; real query traffic can differ, pass it as queries when available
  note: HNSW recall at fixed parameters falls as the corpus grows, so a sweep on a sample is optimistic for HNSW; IVF nlist is set from the row count (4*sqrt(n)), so its recall at a given nprobe can move either way with size. Re-run at production size before fixing parameters
  note: queries_per_second is the median of 3 runs on this machine with one query batch; it ranks settings against each other and is not a serving throughput
  note: latency_ms_p50/p95/p99 time 200 queries one at a time in this process, with no network, concurrency or serving stack; compare settings with them, do not quote them as service latency
  note: index_bytes is the serialized index (vectors plus graph or lists); the process needs more while building, and a product-quantized index would be smaller

Queries default to a seeded sample of the corpus; pass --queries to time real queries, and do so whenever a query log exists: the setting that is fastest at the target recall on corpus-drawn queries can differ from the one real traffic needs.

• Each setting is timed three times and the median kept (--timing-repeats).
• --latency-queries (default 200, 0 to skip) are also searched one at a time for the percentiles.
• index_bytes is the size of the serialized index, so a setting can be judged on memory as well as speed.
• IVF's nlist follows the row count (4*sqrt(n)) unless --nlist fixes it. To tune nprobe for an existing index, pass that index's nlist.

The notes at the end are part of the JSON as caveats. The MCP tool is sweep_index, with the same options as parameters (n_queries, target_recall, device, margin, timing_repeats, latency_queries, nlist).

The sweep answers one question: which parameters reach the target recall on this corpus and these queries, and what they cost. The numbers are relative to this machine and this process; confirm the chosen setting in the serving environment with its own latency, throughput and memory measurements before fixing it.

With a GPU build of faiss (pip install faiss-gpu-cu12 on Linux) the IVF index and the exact baseline run on the GPU automatically; the output header names the device. HNSW has no GPU implementation in faiss and stays on the CPU. Force a device with --device cpu|gpu.

Compare two embeddings

When two models embedded the same items, or one model did before and after a change, vecscope compare tells how far the two spaces agree on who is near whom. It reports:

• how often the nearest neighbor and the top-k lists coincide
• the near-duplicate share in each space and in both
• whether both spaces pick the same duplicate partner
• the rows that disagree most

It carries no verdicts. Its job is to separate item problems from model problems:

• a 10% collapse that both spaces flag with the same partners is a property of the items (the same image under several ids, say)
• duplicates that only one space flags point at that model

vecscope compare clip.npy lens.npy -k 10          # rows must be the same items in the same order

vecscope x.y.z  compare  n=3000 dim=128/128 k=10  neighbors=recomputed/full
  nn1_agreement=0.7050  topk_overlap=0.8537  topk_overlap_p10=0.7000  rows_with_no_overlap=0.0000
  near_dup A=0.0400  B=0.0400  both=0.0400  only_A=0.0000  only_B=0.0000  same_nearest_among_both=1.0000
    the two spaces share the nearest neighbor for 70.5% of rows and 85.4% of their top-10 lists
    on average (0.0% of rows share nothing). Near-duplicates: 4.0% in A, 4.0% in B, 4.0% in both,
    and 100.0% of those agree on the partner. Duplicates flagged by both spaces with the same
    partner are a property of the items; duplicates flagged by one space only point at that model.
  largest disagreements (row: nearest in A / nearest in B, top-k overlap):
    1769: 1340 / 1340  0.50
    1810: 1229 / 1229  0.50

Read neighbor_scope before quoting the agreement.

Above --sample rows (default 100k) a seeded subset is compared, and by default the neighbors are recomputed exactly inside that subset: an item whose true neighbors fell outside the sample is compared on substitutes, which agree more often than the real ones.

On a 2,000-row synthetic pair of spaces, a 500-row sample reports a top-10 overlap of 0.75 where the full corpus gives 0.57. Such a result is marked neighbor_scope: sample, and the explanation says so.

To compare full-corpus rankings on a sample of rows, pass neighbor lists computed on the full corpus in each space, exact or from the production index (an (n x >= k) id file, self ids and negative fillers are dropped):

vecscope compare clip.npy lens.npy -k 10 --neighbors-a clip_knn.npy --neighbors-b lens_knn.npy

The result is then neighbors=supplied/full: the rows may be sampled, their neighbors are not, and the agreement is between the lists you supplied, exact or approximate as they were computed. The MCP tool is compare_embeddings (neighbors_a_path, neighbors_b_path).

Retrieval quality and the fix order

The findings describe the vector space; whether the right documents come back needs labels.

Pass relevance judgments for the queries and the audit adds a retrieval block:

• nDCG@k, recall@k, MRR@k and hit rate of exact cosine search over every row (also when the findings are sampled)
• duplicate_result_share, the share of result slots taken by a near-duplicate of a higher-ranked result

--ranked scores the top-k your production index returned next to exact search, with overlap_with_exact.

--ranked takes the final top-k of whatever serves the queries: an ANN index, a store query with filters, hybrid search or a reranker. The exact baseline is plain cosine over every row, with no filter, keyword score or reranking.

• With those stages in the pipeline, overlap_with_exact measures how far the pipeline departs from vector search, not ANN recall alone.
• vecscope does not score the stages separately. To find the stage that loses quality, pass each stage's output as --ranked in turn.

vecscope audit corpus.npy --queries queries.npy --qrels qrels.tsv
vecscope audit corpus.npy --queries queries.npy --qrels qrels.tsv --ranked prod_topk.npy

qrels holds one of:

• query row, corpus row, grade per line (.tsv/.csv, a header is skipped)
• {"query row": {"corpus row": grade}} in .json
• one object per line in .jsonl

BEIR's query-id/corpus-id/score names work. Rows are positions in the query and corpus files.

The scores average over the queries with at least one relevant (grade > 0) judgment, since nDCG and recall are undefined without one. Queries with none, or with no judgment at all, are counted in unscored_queries and named in the explanation.

With --baseline, the diff lists retrieval_changes next to the finding changes, so a change is judged by the finding it was meant to fix and by retrieval:

  retrieval vs baseline:
    exact   duplicate_result_share   0.0512 -> 0.0000  (-0.0512)
    exact   hit_rate_at_k            1.0000 -> 1.0000  (+0.0000)
    exact   mrr_at_k                 1.0000 -> 1.0000  (+0.0000)
    exact   ndcg_at_k                1.0000 -> 1.0000  (+0.0000)
    exact   recall_at_k              1.0000 -> 1.0000  (+0.0000)

retrieval_comparable is false when the two runs used different qrels or k.

The fix order

Every audit also carries a plan: the non-ok findings grouped by the cause they likely share, in the order to fix them. Each step lists:

• its findings with values
• the common cause to check
• the smallest change that tests it
• the numbers to re-measure
• why it comes at that position

  plan (fix in this order; each step is a hypothesis to test):
    1. metric: metric_mismatch=0.0329 fail
       cause: vector norms vary, so cosine and inner product rank differently
       first: decide the index metric; L2-normalize if cosine is intended
       check: norm_cv near 0 and topk_overlap 1.0 on re-audit
    2. shared_direction: anisotropy=0.926 warn, antihubs=0.061 warn, collapse=0.108 fail
       cause: a large component shared by all vectors (a mean offset or a few high-variance coordinates) that compresses scores and creates hubs
       first: subtract the corpus mean and re-normalize (anisotropy.details shows the expected cosine after centering); if rogue_dimensions fails, z-score dimensions instead. Apply the same transform to queries and rebuild the index
       check: anisotropy and rogue_dimensions below warn_at, hubness.skewness down; retrieval ndcg_at_k not lower than before

That is a synthetic corpus with a shared offset on unnormalized vectors (2,000 rows, 32 dims); the diff above is the same corpus after both steps, against the normalized one. Its 80 labeled queries are easy enough that nDCG stays at 1.0, so only the duplicate share shows the change.

The order follows dependencies between fixes:

• broken inputs (zero vectors, copies) first, because they distort every statistic
• then a dominant coordinate (a rogue dimension decides every distance and with it the norms, outliers and hubs, so z-scoring comes before normalizing)
• then the metric
• then transforms such as centering, which rebuild the neighbor graph
• then the model
• then corpus structure and the index
• the query side last, measured against the corrected corpus

A few findings move to the step whose cause can produce them:

• A near-duplicate share caused by a shared direction (every cosine above the cutoff, no exact copies) goes with the direction, not with deduplication.
• A query_gap next to duplicates goes with them. Near-copies shrink the corpus distance the gap divides by, so it is read again after deduplicating before the queries are blamed.
• Islands next to a near-duplicate group smaller than the island size go with the duplicates, since the copies may be the islands.

The rules are fixed and listed in src/vecscope/plan.py. A step is a hypothesis to test with its re-measure line, not a diagnosis.

Each finding in a step carries placed_because, one of:

• "this group's cause covers it"
• the other finding whose verdict put it there, for example "rogue_dimensions is not ok; one coordinate moves this finding too". This is a grouping by co-occurrence, so check it with the step's re-measure numbers.

What the fixes did to retrieval on real embeddings is in the verification.

Read from a vector store

The path can name a faiss index file or a vector database instead of an embedding file.

• Rows are read in ascending id order, so two reads of unchanged contents give the same matrix.
• The evidence records the source with a fingerprint of what was read.
• Row ids in details.sample_ids come with sample_source_ids, the store's own ids.

pip install "vecscope[qdrant]"     # or [chroma], [milvus], [weaviate], [index] for faiss, [stores] for all
vecscope audit index.faiss --ann-from-source --sample 0
vecscope audit "qdrant://localhost:6333/docs?vector=dense" --save-ids ids.json
vecscope audit "qdrant+local:///data/qdrant?collection=docs"
vecscope audit "chroma+local:///data/chroma?collection=docs"
vecscope audit "milvus+local:///data/milvus.db?collection=docs"
vecscope audit "milvus://milvus.internal:19530/docs?api_key_env=MILVUS_TOKEN"
vecscope audit "weaviate://localhost:8080/Docs?grpc_port=50051"

--ann-from-source queries the store's own index for every stored vector and scores it as --ann, so the recall of the index in production is measured without exporting neighbor lists.

Options go in the query string:

• vector= (named vector, or the Milvus field)
• https=1
• pk= (Milvus primary key)
• api_key_env=NAME (the key is read from that environment variable; keys in the URI itself are not used)
• timeout= seconds per request to a network store (default 60)

Reading is read-only, except that a Milvus collection that is not loaded is loaded for the read and released again.

Secrets stay out of the output:

• The evidence records the URI without user info, and with every query value other than the options above replaced by REDACTED.
• An error from a store client is re-raised with the API key and those values scrubbed.
• Chroma's HTTP client takes no timeout of its own, so timeout= is set on its session after it connects; the connection check the client makes when it is created has none.

The MCP server reads local stores inside its allowed directories, and network stores only when started with --allow-network.

Large corpora: incremental audits

An audit of every row is quadratic in the rows. When a corpus grows by appending or has rows re-embedded in place, --knn-cache keeps the exact neighbor lists between runs and computes only what those rows change.

vecscope audit day1.npy --sample 0 --knn-cache knn.npz    # created: all rows
vecscope audit day2.npy --sample 0 --knn-cache knn.npz    # extended: only the appended rows
vecscope audit day3.npy --sample 0 --knn-cache knn.npz    # updated: changed rows and their neighbors

The cache holds a digest of every row it was computed on. It is:

• reused for the same rows
• extended when rows were appended and the cached rows are unchanged
• updated when some cached rows hold new vectors: those rows, and the rows whose neighbor list contained one of them, are recomputed against every row; every other row only looks at the changed and appended rows
• rebuilt otherwise: the rows to recompute would be more than half of all rows (an update would then cost about what a rebuild does), rows were removed, or dim changed. knn_cache.reason says which.

A change spreads: with k=10 every changed row sits in about ten other lists. Changing 1% of the rows at random recomputes about 10% of them, and changing 5% about half, so an update pays off for small, scattered re-embeddings and for appends.

The neighbor lists equal a full computation except among exactly tied distances, and the findings agree to float32 rounding. In Python, vecscope.incremental.cached_knn_for_audit(x, k, path) returns the lists for audit(x, knn=...).

On 100,000 rows x 768 dims (the machine of the performance table):

• appending 5,000 rows took 11 s to extend the cache and 4 s to audit from it, against 94 s for a full audit
• appending 20,000 rows took 43 s plus 4 s against 123 s

Every verdict matched the full audit and the largest relative value difference was 7e-5. Creating the cache costs about what a full audit does (83 s here).

Re-embedding 1,000 of the 100,000 rows in place recomputed 10,625 rows: 15 s to update the cache and 5 s to audit from it, against 95 s for a full audit, with the same verdicts. With 10,000 rows changed an update would recompute 68,933 rows in 83 s, about as long as the full audit's 89 s, so the cache is rebuilt there instead. Re-run with that rule, 10,000 changed rows rebuilt the cache in 87 s against a full audit's 87 s.

Use with an agent

vecscope ships an MCP server with four read-only tools: audit_embeddings, sweep_index, compare_embeddings and describe_findings. Register it with any MCP client:

pip install "vecscope[mcp]"
claude mcp add vecscope -- vecscope-mcp --allow-dir /data/embeddings     # Claude Code

{ "mcpServers": { "vecscope": { "command": "vecscope-mcp", "args": ["--allow-dir", "/data/embeddings"] } } }

The JSON form goes in .cursor/mcp.json for Cursor and in .mcp.json for a Claude Code project. Codex reads TOML from ~/.codex/config.toml:

[mcp_servers.vecscope]
command = "vecscope-mcp"
args = ["--allow-dir", "/data/embeddings"]

Then ask:

Audit the embeddings in data/chunks.npy before we tune the retriever, and tell me
which finding to fix first.

Tools and responses

• audit_embeddings runs the audit, with the same inputs as the CLI.
• sweep_index runs the index sweep.
• compare_embeddings runs the comparison.
• describe_findings lists every finding with its thresholds, causes and actions.

Responses start with a one-line summary and an audit_id. The summary names the non-ok findings, retrieval nDCG, the first plan step and, on a re-audit, the comparison with the baseline.

Pass that id as baseline when re-auditing after a change. The response then carries baseline_diff (verdict and value changes, regressions, improvements, retrieval changes), as --baseline does in the CLI, without the agent saving the earlier response. The server keeps its last 32 audits in memory; a saved response or --json file works too.

verbose=false drops the details, explanations and actions of findings that are ok. Agents without MCP can run the CLI with --json and read the same evidence.

The agent skill in .agents/skills/vecscope/SKILL.md tells the agent when to run the audit and what to do about each verdict: center or swap the model, deduplicate, re-rank, or rebuild the index. The server also carries those instructions, so an agent that only has the MCP connection follows the same loop.

What the server may read

The tools are read-only, but they read whatever path the agent names, so the server limits where that can be. A path outside the allowed directories (after resolving .. and symlinks) is refused before it is opened.

• By default, only the directory it was started in and below.
• --allow-dir DIR names the directories instead, as in the setup above. It can be repeated; the environment variable VECSCOPE_ALLOWED_DIRS (paths separated by : on Unix, ; on Windows) does the same.
• --allow-any (or VECSCOPE_ALLOW_ANY=1) lifts the restriction, so every file the server process can read is reachable. Use it only on your own machine.
• Network vector stores need --allow-network on top. Read-only does not make reaching them harmless, so grant it only where the server's network access is what you intend.

Each file is read through one handle that is checked after opening: it must be a regular file, and the same file the path resolves to inside the allowed directories. A symlink swapped between the check and the read, a FIFO or a directory is refused.

Local database directories (qdrant+local:// and the like) are checked by path only, because their client opens the files itself.

One request cannot exhaust the server's memory:

• A file larger than --max-read-mb (default 4096; VECSCOPE_MAX_READ_MB; 0 for no limit) is refused before it is read.
• A store stops being read once its vectors pass the same size.
• Files are parsed straight from the checked handle, never copied into memory first. On a 500,000 x 384 .npy file (0.72 GiB) an MCP read peaks at 0.79 GiB, as the CLI does.

Calls in parallel, or with expensive settings, are bounded too:

• The server runs each tool call in its own thread. Audits, sweeps and compares take one of --max-concurrent slots (default 1; VECSCOPE_MAX_CONCURRENT), and further calls wait. An agent that fires three audits at once therefore holds one corpus in memory, not three.
• Settings that multiply a call's cost are capped per call: k and clusters at 1,000, sample_repeats and timing_repeats at 10, n_queries at 100,000 and latency_queries at 10,000. The CLI has no caps.

Each audit, sweep and compare writes one line to the server's stderr, which MCP clients keep in their logs. It names the tool and every file or store it read (credentials redacted), the verdict or error, and the time taken, including any wait for a slot:

vecscope-mcp: audit_embeddings path=/data/run/corpus.npy qrels_path=/data/run/qrels.tsv: fail in 2.1 s

The recent audits that baseline refers to are kept behind a lock. A stdio server serves the one client that started it; agents that run in parallel each start their own server.

Tested clients

The diagnose, change and re-verify loop is tested over the MCP stdio protocol (tests/test_agent_workflow.py), and was run end to end by Claude Code in headless mode (claude -p with the server above).

It audited a corpus with a shared offset, normalized, centered and re-audited as the plan said, and stopped at the remaining step it was not asked to fix. Cursor and Codex use the same protocol and were not run.

docs/agent_case.md records two cases on real SciFact embeddings, each run twice:

• Two ingestion faults: a batch ingested twice and a worker that skipped normalization. From the symptom alone, the agent found both, fixed the corpus and re-audited against its first audit_id. Production nDCG@10 went from 0.100 to 0.713.
• An IVF index searched with nprobe=1: the audit found the embeddings healthy and the production ranking 0.29 nDCG below exact search. The agent swept the index, chose nprobe, and production nDCG@10 went from 0.422 to 0.700 and 0.711.

If the server stops connecting

"vecscope-mcp not found" usually means the Python environment it was installed into has been replaced. This happens on hosted notebooks and clusters that rebuild their environment.

Check with pip show vecscope in the interpreter the client uses and reinstall there. Then register the server through that interpreter, so the error names the right environment:

{ "mcpServers": { "vecscope": { "command": "/path/to/env/bin/python", "args": ["-m", "vecscope.mcp_server"] } } }

Use as a library

import numpy as np
from vecscope import audit

embeddings = np.load("embeddings.npy")    # any (n, dim) float array

ev = audit(embeddings, k=10)
ev.verdict                                # "ok" | "warn" | "fail"
ev.findings[0].value                      # measured value of the first finding
ev.to_dict()                              # the full evidence, same as --json

Use in CI

The JSON output is stable across runs and carries schema_version, so it can be committed as a baseline and diffed on every index rebuild.

--baseline prints each finding's change and exits 1 when any verdict got worse (or when --fail-on trips). The pipeline then fails on a regression, not on a warn that was already there.

vecscope audit embeddings.npy --json > baseline.json          # once, reviewed and committed
vecscope audit embeddings.npy --baseline baseline.json        # on every rebuild

  baseline baseline.json: verdict ok -> fail  (n, dim, k, sampling, thresholds or near-dup cutoff differ: values and verdicts are not directly comparable)
    hubness                1.0454 -> 0.8594  (-0.1860)  ok -> ok
    antihubs               0.0383 -> 0.0105  (-0.0278)  ok -> ok
    anisotropy             0.0105 -> 0.0127  (+0.0022)  ok -> ok
    collapse               0.0000 -> 0.8000  (+0.8000)  ok -> fail  WORSE
    ...
  regressions: collapse

The example diffs healthy.npy against collapsed.npy, which have different row counts, hence the note on the first line; a rebuild of the same corpus with the same settings prints no note.

With --json the diff is attached as baseline_diff; the library call is diff_evidence(new, old). A baseline written by an older schema_version is rejected with a message to re-create it.

Audit every row (--sample 0) for the baseline and the check, so the two runs see the same rows. When both runs pass --queries, the baseline's query distances also give query_shift (see Queries and clusters); it shows up under only_in_new in the diff because the baseline itself had no reference.

Tolerances for drift

A verdict is a coarse signal: a value can drift toward the threshold for months without crossing it. Every change row therefore carries:

• worsened: the value moved in the bad direction
• worsening: by how much, in the metric's units, positive = worse

--tolerances turns that drift into a gate independent of the verdict bands:

# tolerances.json: {"collapse": 0.01, "ann_recall": 0.02, "hubness": 0.3}
vecscope audit embeddings.npy --baseline baseline.json --tolerances tolerances.json --require-comparable

• A finding whose value worsens by more than its tolerance is listed under value_regressions (text: VALUE WORSE by 0.0300 > 0.01) and exits 1 even when its verdict is unchanged.
• Findings without a tolerance are judged by verdict alone.
• A value that improves is never a regression.

The tolerance is a budget per rebuild against the committed baseline, so re-commit the baseline when a change was reviewed and accepted.

The run also reports comparable: false when n, dim, k, sampling, thresholds or the near-duplicate cutoff differ from the baseline.

--require-comparable makes that an error (exit 2) instead of a note. That is the right policy for a gate, since a diff across different settings measures the settings, not the corpus. The library call is diff_evidence(new, old, tolerances={...}).

Which rule failed the run

Two different rules can fail a run, and the output keeps them apart:

• The absolute rule is the verdict against the thresholds (--fail-on fail by default, --fail-on warn to be strict).
• The baseline rules are regressions (a verdict got worse) and value_regressions (a value worsened beyond its tolerance).

The defaults are calibrated on public models, so a domain whose healthy corpus sits at warn or fail on some finding should not be blocked by the absolute rule at all. --fail-on none --baseline baseline.json judges only the change against the reviewed baseline.

Every run ends with a gate block (JSON) or line (text) that names the rule that decided:

  gate: FAIL; verdict fail reaches --fail-on fail; verdict worse than baseline: collapse
  gate: pass (baseline)

Treat the gate as a trend tracker first. Which tolerance is "too much" depends on how the findings relate to retrieval quality on your own queries.

The retrieval quality study shows that relation for public datasets, and calibrating thresholds says how to measure it on yours before blocking merges on it.

CLI options

vecscope audit vectors.npy -k 20                       # neighbors per item (default 10)
vecscope audit vectors.parquet --json                  # JSON evidence
vecscope audit vectors.npy --ann hnsw_neighbors.npy    # score ANN recall against exact k-NN
vecscope audit vectors.npy --queries queries.npy       # add query-side findings
vecscope audit vectors.npy --queries q.npy --qrels qrels.tsv   # add nDCG/recall/MRR of exact search
vecscope audit vectors.npy --queries q.npy --qrels qrels.tsv --ranked topk.npy   # and of the production ranking
vecscope audit index.faiss --ann-from-source --sample 0     # read a faiss index or store; score its own neighbors
vecscope audit "qdrant://host:6333/docs" --save-ids ids.json   # vector store URI; write ids in row order
vecscope audit vectors.npy --sample 0 --knn-cache knn.npz   # reuse / extend / update exact neighbor lists between runs
vecscope audit vectors.npy --clusters 10               # append a per-cluster report (at most n/10 clusters)
vecscope audit vectors.npy --thresholds th.json         # override (warn, fail) pairs
vecscope audit vectors.npy --fail-on warn              # exit 1 on WARN as well
vecscope audit vectors.npy --fail-on none --baseline old.json   # ignore the absolute thresholds; fail only on regressions
vecscope audit vectors.npy --baseline old.json         # diff against an earlier --json run; exit 1 on regression
vecscope audit vectors.npy --baseline old.json --tolerances tol.json   # also exit 1 when a value worsens beyond its tolerance
vecscope audit vectors.npy --baseline old.json --require-comparable    # exit 2 when the baseline used different settings
vecscope audit vectors.npy --near-dup 0.98             # cosine threshold for near-duplicates
vecscope audit vectors.npy --sample 0                  # audit every row (default samples 100k)
vecscope audit vectors.npy --sample-repeats 3          # re-audit 3 samples; report value ranges and verdict stability
vecscope sweep vectors.npy --target-recall 0.95        # HNSW/IVF parameter sweep (faiss extra)
vecscope sweep vectors.npy --device gpu                # IVF on the GPU (faiss GPU build)
vecscope sweep vectors.npy --margin 0.05               # headroom under which the pick is marked tight (default 0.02)
vecscope sweep vectors.npy --n-queries 2000 --timing-repeats 5   # queries to time (default 1000), timing runs per setting (default 3)
vecscope sweep vectors.npy --latency-queries 0         # skip the single-query latency percentiles (default 200 queries)
vecscope sweep vectors.npy --nlist 128                 # sweep nprobe on the production index's list count
vecscope compare a.npy b.npy -k 10                     # agreement between two embeddings of the same rows
vecscope compare a.npy b.npy --neighbors-a a_knn.npy --neighbors-b b_knn.npy   # full-corpus neighbors on a sample of rows
vecscope calibrate fit_*.json --min-quality 0.45 --out th.json   # gate thresholds from your own pipeline states
vecscope calibrate fit_*.json --min-quality 0.45 --holdout hold_*.json --include-low-confidence   # check on held-out states; also write exploratory pairs

Input formats

The embedding file holds one vector per row. Every format is converted to a 2-D float32 array of shape (n, dim) before the audit.

Values must be finite: NaN, Inf and magnitudes above 1e16 (whose squares overflow float32) are rejected with a message.

In the examples embeddings is that array, for instance the output of a model's encode:

embeddings = model.encode(texts)        # shape (n, dim), e.g. (3000, 128)
embeddings[0]                           # the vector of the first item
# array([ 0.0312, -0.1187,  0.0459, ...], dtype=float32)

Extension	Expected layout	Example
.npy	a 2-D numeric array	np.save("vectors.npy", embeddings)
.npz	the first array in the archive is used	np.savez("vectors.npz", emb=embeddings)
.parquet, .pq	a list-of-float column whose name contains emb or vec; other columns (ids, text) are ignored	pd.DataFrame({"id": ids, "embedding": list(embeddings)}).to_parquet("vectors.parquet")
.jsonl	one JSON object per line; the first list-valued field is the vector	{"id": 7, "text": "...", "embedding": [0.1, -0.3, ...]}
.json	a JSON array of objects (as above) or an array of lists	[[0.1, -0.3, ...], [0.2, 0.5, ...]]
.csv, .tsv, .txt	one vector per line, numbers only, no header or id column	0.1,-0.3,0.8,...

How much memory a format takes to load, measured on 50,000 x 384 float32 vectors (a 73 MiB matrix):

Format	How it is read	Peak memory
.npy	as stored	0.10 GiB
.csv	in chunks of lines	0.18 GiB
.jsonl	line by line, converted in blocks of rows	0.34 GiB
.parquet	only the vector column, in batches, into one array	0.39 GiB
.json	the whole file parsed at once	1.13 GiB

For large corpora use .npy, .csv or .parquet. .json is parsed whole and needs several times the matrix size.

The optional inputs:

• --ann (ann_path in MCP): an integer array of shape (n, k). Row i holds the ids of the k neighbors an ANN index returned for item i, in the same row order as the embedding file. Any of the formats above works.
• --queries (queries_path in MCP): query embeddings in any of the same formats, with the same dim as the corpus.
• The query_shift reference: an earlier --json output of an audit run with --queries (--baseline, or in MCP baseline or query_baseline_path).

ANN ids refer to row positions, so the file must come from an index built on exactly these rows, and every row has to be audited. Above the --sample size (100k rows by default) pass --sample 0, or the audit refuses. Ids that point at the item itself are ignored, so an index that returns self first should be queried with k + 1.

The query_shift reference is used only when it was audited on the same corpus or an unchanged prefix of it (checked with corpus_fingerprint, a SHA-256 of the rows). After normalizing or centering, the earlier query distances are in another space, so query_shift is skipped and query_reference_note says why.

Performance

Exact k-NN is quadratic in the number of rows: doubling the rows roughly quadruples the time, and a full audit (--sample 0) of 100k rows takes minutes on a laptop.

The k-NN pass works in row blocks sized to keep its distance matrix near 1 GiB whatever the row count. The other passes need a few float32 copies of the matrix, so the peak grows with rows x dims at roughly four times the matrix size.

Measured on an 8-vCPU AMD EPYC 7R13 with 8 BLAS threads (NumPy only, single process, clustered synthetic corpus, k=10). Memory is the peak the audit itself allocates; the input matrix comes on top. A three-run median in separate processes agreed with these single runs within 2%, and the OS peak memory within 0.04 GiB.

rows	dim	seconds	peak memory (input excluded)
10,000	384	1.3	0.40 GiB
10,000	768	2.0	0.41 GiB
10,000	1,536	3.6	0.44 GiB
50,000	384	17	1.63 GiB
50,000	768	23	1.70 GiB
50,000	1,536	36	1.84 GiB
100,000	384	65	1.71 GiB
100,000	768	86	1.85 GiB
100,000	1,536	130	3.45 GiB
200,000	768	338	3.46 GiB
200,000, default 100k sample	768	85	2.14 GiB

• Time grows close to n² (3.5 to 3.9x per doubling from 50k up).
• Memory is flat from 50k rows at 384 and 768 dims because the k-NN block shrinks with n, and steps up at 1,536 dims and at 200k rows.
• The default 100k-row sample keeps any corpus near the 100k rows.
• --queries adds one more pass over the corpus per query batch, and metric_mismatch adds a second k-NN pass only when vector norms vary.

To re-measure on your machine at the sizes you care about:

python examples/benchmark_scale.py --rows 50000 100000 --dim 768   # the audit table above
python examples/benchmark_sweep.py                                 # the index sweep

Both take --corpus (clustered, isotropic, anisotropic, duplicates, lowrank, zipf) and record the machine in their JSON.

Sampled audits

Blocking bounds memory, not work: the exact pass still costs about n² × dim multiply-adds, so a million rows at 768 dims is roughly 100 times the 100k-row figure.

For corpora that size, audit the default sample (or several with --sample-repeats) and reserve --sample 0 for the findings that need every row. A sampled audit is a statement about the sample. Which findings need every row:

finding	what a sample gives	needs every row
anisotropy, effective_rank, rogue_dimensions, norms, metric_mismatch, distance_concentration	a distribution estimate: within 2.3% of the full value on 200k x 384 with 100k, 50k and 20k samples, and the verdict holds	no
hubness, antihubs	a lower bound: k-occurrence concentrates with density, so the same corpus scores less hubness on a smaller sample (skewness 2.6 warn on all 200k rows, 1.15 on a quarter, 0.16 on a tenth). Verdicts other than fail are marked inconclusive	to measure the corpus's own hubness
collapse	the duplicate share seen inside the sample: a pair counts only when both rows were drawn, so the sampled share is 0.96 to 1.03 x (sample fraction x full share); the verdict uses the estimated full share, at least the exact copies counted over every row	to count or list every near-duplicate
outliers	an estimate that over-counts on small samples (3.9x at a tenth); verdicts other than ok are marked inconclusive	to find every outlier
connectivity	the islands of the sample's own k-NN graph, which miss most islands of the corpus (37% of rows in islands scored 0.08% on a quarter sample); verdicts other than fail are marked inconclusive	yes, to judge connectivity at all
ann_recall	not available on a sample: ids must line up with the ANN file	yes (--sample 0)
query_gap, query_shift, query_hubness, ood_queries	reference distances from the sample; every query is scored. query_gap is a ratio of two distances that both grow on a smaller sample; compare query_shift only against a baseline with the same sample size	no

A sampled audit therefore:

• lists inconclusive_findings (the verdicts the table marks inconclusive)
• flags them INCONCLUSIVE (sampled) in the text output and names them in the gate line
• takes the collapse verdict from details.estimated_full_share (see below)
• repeats each caveat in the finding's explanation next to the number

estimated_full_share is the sampled share divided by the sample fraction, or the share of rows with an exact copy counted over every row (details.full_corpus_exact_copy_share) when that is larger. Counting exact copies over the whole corpus is cheap (one digest per row), so rare copies the sample missed still count. Near-duplicates that are not exact copies are only seen in the sample.

The 200k-row figures come from the verification; tests/test_regression_injection.py checks the same properties at 6k rows on every run.

Every number is exact for the rows it saw; vecscope never substitutes an approximate neighbor search for the exact one. The only approximation is the choice of rows.

How vecscope works

vecscope computes exact k-NN once over the matrix and derives every diagnostic from it:

• the k-occurrence distribution gives hubness and anti-hubs
• nearest-neighbor cosines give collapse, outliers and distance concentration
• the k-NN graph gives connectivity
• random pairs and the covariance spectrum give anisotropy, effective rank and rogue dimensions
• vector norms give the norm spread
• a second inner-product pass (skipped when norms are uniform) gives metric mismatch
• the ANN neighbor file, if present, is compared against the exact neighbors for recall

Query embeddings, if present, get their own exact search against the corpus for the query-side findings.

Everything is deterministic: fixed seeds, exact neighbors, no learned components, no network. Two runs on the same file give the same JSON.

Exact k-NN is quadratic in the number of rows, so inputs above 100,000 rows are audited on a seeded random sample unless you pass --sample 0.

Real embeddings

Thresholds are calibrated on public, pre-computed embeddings from seven models over eight datasets: OpenAI ada-002 and text-embedding-3-large, Cohere embed-v3 (Korean and English Wikipedia), all-MiniLM-L6-v2, bge-m3, mxbai-embed-large-v1 and gte-base-en-v1.5.

The table shows the measured value for the first 40k rows of each dataset; cells without a label are ok. python examples/real_models.py reproduces it, no model download needed.

finding	Cohere v3 (ko wiki)	Cohere v3 (en wiki)	ada-002 (dbpedia)	3-large (dbpedia)	MiniLM-L6 (wiki)	bge-m3 (wiki)	mxbai-large (wiki)	gte-base (arxiv)
hubness	1.47	1.16	1.56	1.21	1.59	1.85	1.43	2.28 warn
antihubs	0.009	0.005	0.013	0.005	0.022	0.026	0.019	0.013
anisotropy	0.333	0.228	0.692	0.086	0.026	0.307	0.330	0.210
effective_rank	0.158	0.178	0.090 warn	0.148	0.285	0.142	0.087 warn	0.136
collapse	0.002	0.002	0.000	0.000	0.001	0.000	0.017	0.000
rogue_dimensions	3.2	3.2	3.0	2.5	1.7	1.8	2.0	1.8
norms	0.000	0.000	0.000	0.000	0.000	0.000	0.000	0.007
metric_mismatch	1.000	1.000	1.000	1.000	1.000	1.000	1.000	0.952
distance_concentration	1.72	1.74	1.74	1.43	1.78	1.84	2.21	1.62
outliers	0.001	0.001	0.000	0.000	0.001	0.002	0.003	0.000
connectivity	0.000	0.000	0.000	0.000	0.000	0.000	0.000	0.000

Re-running the audit on the first 40k rows of each set reproduces all 88 verdicts. The only warns are hubness on gte-base and effective_rank on ada-002 and mxbai-large.

ada-002's well-known anisotropy (0.692) is below the warn threshold of 0.85: its mean unit vector has norm 0.83, so unrelated documents start from a cosine floor near 0.7. Its successor is isotropic on the same corpus (norm 0.29). Centering brings every model here to a random-pair cosine of 0.00, which is why the finding reports that number.

Hubness skewness of 1.2 to 1.9 is the normal range for text embeddings, so the warn threshold sits at 2.0; gte-base on arXiv abstracts sits just above it. The anisotropy thresholds (0.85, 0.95) come from the retrieval-quality study below.

Do the findings predict retrieval quality?

Partly, and in a specific way. docs/retrieval_quality.md measures every finding against nDCG@10 on two labeled BEIR datasets (SciFact, NFCorpus) with exact cosine search, for eight public models and for controlled defects injected into the best one. python examples/retrieval_quality.py reproduces it.

• Within one pipeline the findings track damage. Injecting a rogue dimension, cutting the rank, adding noise or duplicating rows moves the watched finding to fail (from ok, or from warn where the corpus already carried a few duplicates) while nDCG@10 falls, and the drop is monotone in the defect level.

  Queries embedded with an unrelated model score query_gap 8 to 10 (fail, nDCG@10 near 0); bge-base queries on the gte-base corpus, which share a space, score 2.9 to 3.9 and keep nDCG@10.

  Heavy near-duplication also raises query_gap, because the copies shrink the corpus distance it divides by; the plan lists it with the duplicates so it is read again after deduplicating.

• Across good models the geometry findings do not rank the models. The best models on both datasets (gte-base, bge, e5) have the highest anisotropy (0.58 to 0.82, counting the 19 models of the verification) and ordinary effective rank.

  What does separate the weak models is norms with metric_mismatch (unnormalized outputs) and hubness, at Spearman |rho| of 0.4 to 0.8 with nDCG@10.

• Anisotropy alone does not hurt ranking. Adding a shared offset until the random-pair cosine reaches 0.98 leaves nDCG@10 unchanged. It compresses scores, so absolute cosine thresholds and score calibration break; distance_concentration is the finding that moves when the signal itself is lost.

The anisotropy, query_hubness, query_gap and query_shift thresholds come from the verification: 19 models and injected query faults on three labeled datasets, re-checked on three more.

What it does not cover is how well warn and fail predict a quality drop in other domains or on production query logs. Until that is measured for your domain, read a verdict as a reason to look and a baseline regression as a reason to compare retrieval, not as a quality score.

Calibrating thresholds to your domain

The defaults are calibrated on public models and public datasets; a fail on your corpus is a reason to look, not proof that retrieval is bad, and all-ok is not proof that the right documents come back.

Before using a threshold or a tolerance as a blocking gate, measure what it predicts on your own data:

1. Collect labeled queries from the domain (a few hundred with relevant document ids is enough) and split them: one part to pick thresholds, a held-out part to check them.
2. Run vecscope audit --json on each candidate state of the pipeline (models, chunkers, transforms, index settings) and compute retrieval quality (nDCG@k, Recall@k, MRR) with the same queries on the same state. examples/retrieval_quality.py does both for the public datasets and can be pointed at your own.
3. Relate each finding's value to the quality metric across those states. A finding that does not move with quality in your domain should not gate it, whatever its default says.
4. Set --thresholds and --tolerances from the states where quality was acceptable, using only the first split.
5. Confirm on the held-out split that the thresholds separate the acceptable states from the rest. A threshold tuned and checked on the same queries will look better than it is.

vecscope calibrate does steps 3 to 5 from the audits of step 2. Audit each state with the fitting queries, and each again with the held-out queries:

vecscope audit state_a.npy --queries q_fit.npy --qrels qrels_fit.tsv --json > fit_a.json
vecscope audit state_a.npy --queries q_hold.npy --qrels qrels_hold.tsv --json > hold_a.json
# ... the same for every state, then:
vecscope calibrate fit_*.json --min-quality 0.45 --holdout hold_*.json --out th.json
vecscope audit vectors.npy --thresholds th.json

For each finding it reports:

• the Spearman correlation with the retrieval metric across the states (--metric, nDCG@k by default), its 95% bootstrap interval, and whether the sign is the one the finding implies
• a one-sided permutation p-value in that direction, Holm-adjusted over every finding tested
• the worst value among acceptable states and the best among the others
• a (warn, fail) pair only when the finding tracks quality and separates the two groups, placed a third and two thirds of the way across the gap

A finding that does not move with quality on your data, or moves the other way, gets no pair: keep its default as a reason to look.

With about fifteen findings tested at once, one of them tracks quality by chance on a few states. So a pair goes into --out only when its adjusted p-value is below 0.05:

• The other pairs are reported as exploratory and left out of the gate file. --include-low-confidence writes them too.
• On five states nothing can reach that bar (the smallest one-sided p-value is 1/120 before the adjustment). Ten or more states are recommended, and the report says when there are fewer.

Two checks judge the proposal on states it was not fitted on:

• leave_one_out fits on all states but one and judges the one left out, for every state.
• The holdout line counts the unacceptable --holdout states the gate catches and the false alarms on acceptable ones. Holdout states from another dataset check whether the thresholds carry over.

Calibration needs at least three states, with some on each side of --min-quality. States scored on different qrels or k are refused, because their quality values do not compare (--allow-mixed-qrels overrides).

The same procedure applies to a suggested_action: it is a hypothesis about the cause, and centering, standardizing or truncating vectors changes rankings.

• Keep the original vectors and retrieval results as the baseline.
• Apply the transform to a copy.
• Measure quality, latency and memory on both.
• Adopt the change only when the labeled queries improve, not when the finding turns ok.

FAQ

Does it send my vectors anywhere?

No. vecscope is NumPy only and runs entirely on your machine. It never loads a model or opens a network connection.

How large a matrix can it handle?

Exact k-NN is quadratic in the number of rows. By default vecscope audits a seeded random sample of 100,000 rows when the input is larger, and says so in the output (sampled: true, n_total).

A sampled result describes the sample, not the whole corpus: rare clusters, small islands and a handful of outliers can be missed.

Row ids in details (sample_ids, sample_pairs, the cluster report) always refer to rows of the input file, also when sampled. Use --sample 0 to audit everything; see performance for what that costs.

To see how much the sample itself moves a number, pass --sample-repeats 3 (MCP: sample_repeats). vecscope then audits three independent seeded samples and attaches value_range (minimum to maximum across samples) and verdict_stable to every finding, and lists unstable_findings at the top level.

A finding whose verdict flips between samples should be read as borderline, not as a property of the corpus. The cost is three audits.

value_range is the spread between samples, not an interval that contains the full-corpus value: hubness is a lower bound on every sample, and even a very stable finding can sit a little outside the range of three samples.

Not every finding means the same thing on a sample: hubness is understated, outliers overcounted, islands mostly invisible. The performance table says what each finding can stand for and which verdicts a sampled audit marks inconclusive.

My production model fails the anisotropy check. Is the model broken?

Not necessarily. Many popular text embedding models are strongly anisotropic by construction.

The finding tells you that raw cosine scores are compressed into a narrow band; centering or whitening the vectors, or re-ranking, usually restores discrimination. The details block reports the cosine after centering so you can see whether that is enough. The verdict is evidence for the agent, not a grade for the model.

Which agents can use it?

Any agent that supports local MCP servers, and any agent that can run a shell command and read JSON. The skill file describes both paths.

Documentation

• Agent skill: when to run the audit and what to do with each verdict
• Retrieval-quality study: every finding against nDCG@10 on labeled data
• Verification: exactness, recall scoring, the sampling claims, the query findings, interventions, agent integration, and 19 models on 9 labeled datasets
• Agent cases: an agent finding and fixing two ingestion faults, and a mis-tuned IVF index, over MCP, with the tool calls and numbers
• Examples: synthetic matrices, the calibration script and the retrieval-quality study
• Changelog

Contributing

Issues and pull requests are welcome. See CONTRIBUTING.md for how to run the tests and what a good change looks like. Security reports go through SECURITY.md.

License

MIT