Hybrid search and cross-encoder reranking

Two more changes on top of the rebuilt ingestion. BM25 fused with the dense retriever by reciprocal rank fusion, because alphanumeric model codes such as LP-512-ENG, KAB-545-15 and 6M125 fragment into subword pieces that carry almost no embedding signal. Then a cross-encoder over the fused candidates, which sees the query and the passage together and can therefore tell "125 CFM diesel Davey" from "125 CFM gasoline Ingersoll-Rand".

How often should the compressor oil filter be cleaned on the 125 CFM 100 PSI skid-mounted gasoline compressor?

Correct source: tm-5-4310-221-15 · known wrong-variant sources: tm-5-4310-354-14

Baseline

baseline · 15 ms · relevance 0-1

1	doe-sourcebook p.61	0.74
2	tm-3-4310-100-20p-altscan p.16	0.73
3	tm-5-4310-339-15 p.21	0.73
4	tm-3-4310-100-20p-altscan p.15	0.72
5	tm-9-4310-227-15 p.37	0.72

+ structure

structure · 22 ms · relevance 0-1

1	tm-5-4310-339-15 p.21 PRESSURE LIMITS, PSIG › COM PONENT ACTION IN	0.75
2	tm-5-4310-221-15 p.26 ITEM ITEM TO BE CHECK FOR AND HAVE REPAIRED	0.74
3	tm-3-4310-100-20p p.13 DRAIN PLUG › ™~® AR601106	0.73
4	tm-3-4310-100-20p-altscan p.15 j \ OIL STRAINER › DRAIN PLUG	0.73
5	tm-3-4310-100-20p p.22 2-18. Compressor Unit, Model AN-M4D › 2-19.	0.72

+ hybrid

structure+hybrid · 21 ms · relevance 0-1

1	tm-5-4310-221-15 p.192 3. Fire Protection › 4. Lubrication	0.71
2	tm-3-4310-100-20p p.22 2-18. Compressor Unit, Model AN-M4D › 2-19.	0.72
3	tm-9-4310-227-15 p.2 Depot Maintenance Manual › 15 CFM, 175 PSI;	0.69
4	tm-5-4310-221-15 p.112 89. Engine Oil Filter › a. Removal. Remove t	0.69
5	tm-3-4310-100-20p-altscan p.16 CANVAS COVER, BACK REST HARNESS › AND STRAP	0.69

+ rerank

improved · 2570 ms · relevance 0-1

1	tm-5-4310-221-15 (was #7) p.26 ITEM ITEM TO BE CHECK FOR AND HAVE REPAIRED	0.99
2	tm-5-4310-221-15 (was #1) p.192 3. Fire Protection › 4. Lubrication	0.98
3	tm-5-4310-221-15 (was #46) p.34 AGO 9048A › 4. Identifi.=*ion and Tabulated	0.94
4	tm-5-4310-221-15 (was #13) p.29 OPERATOR, ORGANIZATIONAL, FIELD, AND DEPOT M	0.94
5	tm-5-4310-221-15 (was #8) p.21 Operator, Organizational, Direct and General	0.91

The reranker is not the whole fix A cross-encoder reorders what it is given; it cannot recover a passage that never entered the candidate pool, and it has no idea which revision is current. The revision and variant guards below run after it and are plain code — a rule that says a superseded edition loses to a current one is not something a language model should be asked to remember.

Final ranking, with every adjustment shown

[1] TM 5-4310-221-15 — Ingersoll-Rand GER-125, rotary, skid mounted, gasoline, 125 CFM 100 PSI

page 26 · ITEM ITEM TO BE CHECK FOR AND HAVE REPAIRED NOT READY/ › NO. IWIMIH INSPECTED OR ADJUSTED AS NECESSARY AVAILABLE IF: · May 1963 · Ingersoll-Rand · GER-125

Family Ccorrect machine

ITEM ITEM TO BE CHECK FOR AND HAVE REPAIRED NOT READY/ › NO. IWIMIH INSPECTED OR ADJUSTED AS NECESSARY AVAILABLE IF:
9 300} Engine crank- Drain and refill engine oi] case every 300 hours or 6 months. 10 300} Engine oi] fil- |Replace engine oil filter ter every 300 hours or 6 months (para 89). 1] 300} Air receiver Drain and refill the compres- oil separator sor lubricating system every 300 hours or 6 months. 12 300} Compressor oi] |Remove and clean compressor filter oil filter every 300 hours or 6 months (para 98).
Figure 13. Organizational Preventive Maintenance Checks and Services. (Sheet 2 of 2)
TH 5-4310-221-15 C5 By Order of the Secretary of the Army:

relevance 0.99 · dense cosine 0.745 · cross-encoder logit +1.26 · ranking score 4.256