Executed after the canonicalization pass (see
canonicalization_20260419.md) to lift the 15 real entity_ids from
their Similar-50 cosine baseline (~0.30 enrichment_score, sparse
metadata, no fresh embeddings) to TandemStride-style density on the
/companies/[slug] pages.
Run was done locally on the user's M-series Mac via llama.cpp — the Linode GPU burst cluster was unreachable at run time, and the targeted 15-entity scope is small enough to complete in minutes on the local GPU without needing a provisioned cluster.
Infrastructure used
- Inference:
llama-serveron:8080, modelgemma-4-E4B-it-Q4_K_M.gguf, 16k context,--n-gpu-layers 99. Launched in a detachedscreensession. - Embedding:
llama-server --embeddingon:8081, modeljina-embeddings-v5-text-small-retrieval/v5-small-retrieval-Q8_0.gguf, 8k context,--n-gpu-layers 99. Produces 1024-dim vectors (matchesVECTOR_DIMinconfig.py). - Pipeline:
enrich_cli.py entity <id>for core enrichment (Q0 grounding → Q1 profile → Q2 media → Q3 relationships → Q4 character stats → Q5 visual profile → taxonomy), followed by a directget_embeddings_batchpass through the shared_build_entity_embed_texthelper for deterministic re-embed.
Sequence
- Smoke-run on
#737992(UnitedHealth Group) with 4k ctx:Q1 core profilefailed withrequest (4184 tokens) exceeds the available context size (4096 tokens). Score landed at 0.972 anyway via Q2–Q5. - Restarted
llama-serverwith 16k ctx. - Batched all 15 entities through
enrich_cli.py entitywith an 8-second pause every 3 entities (per the user's API-rate-limit rule). - Brought up the embedding server on
:8081. - Targeted re-embed of all 15 via Jina-v5 (1024-dim).
- Re-ran enrichment on the five low-scorers (≤65%) with the full stack healthy to give them every possible lift.
Before → After
All 15 went from Similar-50 baseline (~0.30, no fresh embedding,
cosine-peer metadata only) to the scores below. Mean enrichment
score: 0.839. Re-embed coverage: 15/15 with
embedding_updated_at populated.
entity_id score passes re-emb Name
--------- ----- ------ ------ ---------------------------------
737992 97.7% 3 yes UnitedHealth Group Incorporated
1166349 64.0% 4 yes Optum Care
427873 89.7% 4 yes OptumRx
1166348 61.4% 4 yes Optum Bank
833848 100.0% 6 yes OptimizeRx Corporation
629634 100.0% 3 yes Universal Health Services
673584 65.2% 4 yes Sharp HealthCare
673545 65.3% 3 yes Prime Healthcare
573586 72.0% 3 yes Deloitte Canada
1317492 95.6% 3 yes Deloitte Risk & Financial Advisory
825797 100.0% 2 yes Marsh McLennan
827137 61.0% 3 yes MagnaChip Semiconductor Corp.
856177 100.0% 6 yes Skyworks Solutions
387081 93.5% 2 yes SMIC
389485 93.5% 2 yes Meta Platforms Inc.
Score-ceiling finding
Five entities plateaued at 61–65% across multiple passes even with the full pipeline (16k ctx inference + Jina embed + Q0 web grounding):
- Optum Care · Optum Bank · Sharp HealthCare · Prime Healthcare · MagnaChip Semiconductor
The ceiling is public data availability, not pipeline health:
- Q0 grounding returns thin snippets for these entities because they're subsidiaries / specialty operators / less-covered niches.
- Q5 visual-profile readiness sits in the 0.60–0.64 range for each.
- Taxonomy bonus caps at 0.10–0.18.
This validates the burst runbook's Prerequisite #2 — authoring a
per-company seed_company_deep/data/<slug>.py CompanySpec with
hand-curated facts is the only way to push these five above 0.70
without misattributing generic industry data to the company.
Pre-existing bugs surfaced
enrichment_model+enrichment_hostcolumns are still NULL on every row after enrichment. Same bug called out in the 10× fleet debrief (0YFNFkQTe3hptNBSkTqski). Fix is inenrich.py, not scoped here.
Impact on /companies/[slug]
Every one of the 15 newly-canonical /companies/[slug] pages now
renders the shared OrgStatisticsBlock with real enrichment data
(via the empty-state bypass added in PR #231), plus fresh 1024-dim
Jina embeddings driving cosine-similar lookups elsewhere in the
graph. No visible change needed to the frontend.
Infra follow-up
Local screen sessions llama-server (port 8080) and llama-embed
(port 8081) are still running. To stop both:
screen -S llama-server -X quit
screen -S llama-embed -X quit
pkill -f llama-server
Or leave them up if you expect to re-run enrichment soon.