Running today
Real output from the development system, October 2026. The index holds 2,558 vehicles: 2,550 from the
public EPA fuel-economy dataset (model years 2024–2025) enriched with NHTSA crash ratings and recalls,
plus 8 hand-curated vehicles. Automotive service documentation is being added next.
Fleet-scale ingestion
Measured
Vehicles are chunked, embedded on the GPU and written to a two-tier index (vehicle summaries for
routing, detailed chunks for answers) in checkpointed batches of 50 that are safe to re-run.
29 / s
Vehicles ingested, single stream
86.6%
Golden-set recall at this scale
Numeric questions need structure
Before / after
At fleet scale, text search matched the number in the question to numbers in model names. Typed
attributes now let threshold and superlative questions run as SQL, while other questions stay on
hybrid text search.
"which vehicles have over 500 horsepower"
// Before: hybrid text search, top results
mercedes-benz-s-class-2025-s500-4matic-... // "500" in the trim name
lexus-ls-2025-500-rwd-3-4l-6-cylinder-... // same
0 of 3 expected vehicles returned
// After: routed to structured search
SELECT ... FROM vehicle_attributes
WHERE horsepower > 500 ORDER BY horsepower DESC
3 of 3 expected vehicles returned: BMW M3 Competition, Chevrolet Corvette Z06, Rivian R1T
Local inference on NVIDIA hardware
RTX 5090
The model server and GPU-accelerated embedding and reranking services run on a local RTX 5090. Nearly
all 32 GB of VRAM is held by resident model weights and KV cache. The serving processes run in
containers, so they don't appear by name in the Windows process list.
nvidia-smi
PS> nvidia-smi
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 610.88 KMD Version: 610.88 CUDA UMD Version: 13.3 |
+-----------------------------------------+------------------------+----------------------+
| GPU Name Driver-Model | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
|=========================================+========================+======================|
| 0 NVIDIA GeForce RTX 5090 WDDM | 00000000:41:00.0 On | N/A |
| 0% 31C P5 45W / 600W | 31977MiB / 32607MiB | 0% Default |
+-----------------------------------------+------------------------+----------------------+
Measured retrieval quality
Golden-set eval
Retrieval is scored against a 29-query golden set across eight categories and four difficulty levels.
For attribute questions with many correct answers (“electric vehicles”), every vehicle that meets the
criteria counts as relevant, with the relevant sets generated from the data.
| Stage | Passed | Recall | Precision |
|---|
| 8 curated vehicles | 28 / 29 | 98.9% | 53% |
| 2,558 vehicles, original labels | 15 / 29 | 51.7% | 10% |
| Open-world relevance labels | 20 / 29 | 67.6% | 26% |
| Structured search for numeric questions | 26 / 29 | 86.6% | 43% |
| Variant collapsing (one result per model) | 26 / 29 | 86.3% | 44% |
| Category (structured search) |
Passed |
Recall |
|
Precision |
| Features | 2 / 2 | 100% |
| 58% |
| Safety | 3 / 3 | 100% |
| 39% |
| Rankings | 4 / 4 | 100% |
| 15% |
| Performance | 3 / 3 | 100% |
| 13% |
| Fleet filter | 6 / 6 | 92% |
| 92% |
| Maintenance | 3 / 3 | 83% |
| 42% |
| Ownership cost | 3 / 4 | 75% |
| 31% |
| Cross-chunk | 2 / 4 | 52% |
| 29% |
The 3 remaining failures, reported rather than tuned away. Two are subjective questions
("good family car with low running costs" and "best car for someone who loves track days but needs daily practicality")
whose answers aren’t in spec data; query rewriting into structured constraints is the next step. One
is a data gap: only one vehicle has 5-year ownership-cost data.
