← Research archive
Technical note verified Sep 4, 2026

Chroma Corpus Compatibility and Retrieval Equivalence

KESPA validated that the isolated 423-claim clean corpus could be reproduced in a Brain-compatible Chroma layout without changing IDs, documents, metadata, or observed retrieval ordering. All 423 records matched structurally, 32/32 top-1/top-10 retrieval checks matched, and only float-scale embedding round-trip differences were observed.

chroma retrieval compatibility deterministic-validation verified-knowledge
MARKDOWN

README.md

2,129 bytes SHA-256 15c6942d04b11a92…

KESPA-NOTE-008 — Chroma Corpus Compatibility and Retrieval Equivalence

Date: 2026-09-04 Status: Verified — compatibility validation passed Project: KESPA AI / NexLabs Studios

Purpose

Before the clean verified corpus could be used for retrieval calibration and the A/B/C benchmark, KESPA needed to prove that moving the corpus into a Brain-compatible Chroma layout did not silently change the knowledge records or the retrieval behavior.

The validated corpus contained 423 verified atomic claims.

Structural comparison

The compatibility check recorded:

  • IDs exact: 423/423
  • documents exact: 423/423
  • metadata exact: 423/423

Combined structural result:

423/423 exact

That meant the isolated benchmark index represented the same claim records, not a rewritten or transformed copy with materially different metadata.

Retrieval-equivalence check

KESPA then compared retrieval behavior across the compatible representations.

Recorded result:

  • top-1 / top-10 retrieval checks: 32/32 matched
  • result: PASS

This supported treating the isolated Brain-compatible Chroma collection as a valid retrieval substrate for the later policy calibration and benchmark work.

Embedding round-trip differences

The embeddings were not perfectly byte-identical after storage/round-trip:

  • 43/423 embeddings had a nonzero numeric delta
  • maximum absolute delta: 1.49011611938e-08

Those differences were at float-roundtrip scale.

Critically, they did not produce a mismatch in the 32 tested top-1/top-10 retrieval comparisons.

Why this mattered

Without this validation, a later benchmark could be confounded by the index-conversion process itself.

The compatibility test established that:

clean verified corpus -> Brain-compatible Chroma

preserved the tested record and retrieval semantics closely enough for controlled evaluation.

Boundary

This validation did not modify:

  • production Chroma;
  • Brain API code;
  • the production database;
  • knowledge trust assignments.

It was an isolated compatibility check, not a production migration.

TEXT

SHA256SUMS.txt

393 bytes SHA-256 775da46a6cc908d7… 
15c6942d04b11a92a0daf0863284f6d4f48d673bc001ac1decdebcb47d3875fa  README.md
0f85ea884aef161cc684038eb5e507f76822be1a606874bf1ecade1899aa1b5f  note.json
3573d0e465610c4a16235509a1f33a14a7a6a91fdf6b17a228c9f0cc8dad81cf  metrics.csv
a80f0a7013efa6a32e1bf09834cc830a17f3267f406cc56cef952a1c06d28239  methodology.md
7b4cc184ca42a81ec4ff7d9f601f9f885475f2e530d680f52307a10e7e503496  provenance.json
MARKDOWN

methodology.md

918 bytes SHA-256 a80f0a7013efa6a3…

Methodology

Corpus under test

The test used the clean 423-claim verified corpus and a Brain-compatible Chroma representation using all-MiniLM-L6-v2 embeddings with cosine distance.

Structural validation

IDs, documents, and metadata were compared across all 423 records.

The validator required exact agreement for those fields.

Retrieval validation

A fixed retrieval-equivalence sample compared top-1 and top-10 results across the compatible index representations.

All 32 recorded checks matched.

Embedding comparison

Stored embeddings were compared numerically.

Tiny float-roundtrip differences were permitted only when they remained small and did not alter the tested retrieval results.

The maximum recorded absolute delta was 1.49011611938e-08.

Isolation

The work was performed outside the production Chroma path and did not modify Brain API behavior or production database state.

CSV

metrics.csv

416 bytes SHA-256 3573d0e465610c4a… 
metric,value,unit_or_status
verified_claims,423,claims
ids_exact,423/423,records
documents_exact,423/423,records
metadata_exact,423/423,records
top1_top10_retrieval_matches,32/32,checks
embeddings_with_float_delta,43/423,vectors
max_embedding_delta,1.49011611938e-08,absolute_delta
retrieval_equivalence,PASS,status
production_chroma_modified,NO,boundary
brain_api_modified,NO,boundary
database_modified,NO,boundary
JSON

note.json

2,086 bytes SHA-256 0f85ea884aef161c… 
{
    "schema": "kespa.public_technical_note.v1",
    "id": "KESPA-NOTE-008",
    "title": "Chroma Corpus Compatibility and Retrieval Equivalence",
    "date": "2026-09-04",
    "status": "verified",
    "purpose": "Document the compatibility validation performed before using the 423-claim clean verified corpus in Brain-compatible Chroma retrieval experiments.",
    "corpus": {
        "verified_claims": 423,
        "embedding_model": "all-MiniLM-L6-v2",
        "distance_metric": "cosine"
    },
    "structural_equivalence": {
        "ids_exact": "423/423",
        "documents_exact": "423/423",
        "metadata_exact": "423/423",
        "combined_ids_documents_metadata_exact": "423/423"
    },
    "retrieval_equivalence": {
        "top1_and_top10_checks": "32/32",
        "result": "PASS"
    },
    "embedding_roundtrip": {
        "embeddings_with_nonzero_float_delta": "43/423",
        "maximum_absolute_delta": 1.49011611938e-8,
        "interpretation": "Float-roundtrip scale only; no observed retrieval-equivalence failure."
    },
    "research_boundary": {
        "production_chroma_modified": false,
        "brain_api_modified": false,
        "database_modified": false,
        "trust_assignment_changed": false
    },
    "interpretation": [
        "The clean verified corpus could be represented in the Brain-compatible Chroma layout without losing record identity, text, or metadata.",
        "Observed float-scale embedding differences did not change the tested top-1/top-10 retrieval results.",
        "The compatibility result supported using the isolated clean Chroma collection for later retrieval calibration and A/B/C benchmarking."
    ],
    "limitations": [
        "The retrieval-equivalence sample comprised 32 checks, not an exhaustive query set.",
        "Embedding equality was not byte-for-byte for every vector because 43 of 423 vectors showed tiny float round-trip differences.",
        "This validation established compatibility and retrieval equivalence under the tested conditions; it was not itself an answer-quality benchmark.",
        "The result is specific to the tested corpus, embedding model, metadata contract, and Chroma configuration."
    ]
}
JSON

provenance.json

902 bytes SHA-256 7b4cc184ca42a81e… 
{
    "schema": "kespa.public_provenance.v1",
    "research_id": "KESPA-NOTE-008",
    "source_basis": "Historical KESPA changelog/Jira export supplied for public research reconstruction.",
    "source_observations": [
        "The clean 423-claim corpus was validated in a Brain-compatible Chroma representation.",
        "IDs, documents, and metadata matched exactly for 423/423 records.",
        "Top-1/top-10 retrieval-equivalence checks matched 32/32.",
        "43/423 embeddings showed only tiny float-roundtrip differences, with maximum absolute delta 1.49011611938e-08.",
        "Production Chroma, Brain API, and database state were not modified by the compatibility validation."
    ],
    "publication_note": "This public note publishes aggregate compatibility measurements and methodology. Raw Chroma storage files, full embeddings, private implementation code, and internal validation scripts are not republished."
}