171
A RAG-based LLM App oach o Da a Valida ion and Ha moniza ion in
Ship Design
Janica B onson, NTNU, Ålesund/No way, Janica.a.b onson@n nu.no
Ma yam Teimou i, UTU, Tu ku/Finland, m ebad@u u. i
Hen ique Gaspa , NTNU, Ålesund/No way, hen ique[email p o ec ed]
Ica o Fonseca, NTNU, Ålesund/No way, ica o. onseca@n nu.no
Ka olina Bie kowska, NTNU, Ålesund/No way, ka olina.bie kow[email p o ec ed]
Filip Gin e , UTU, Tu ku/Finland, ig[email p o ec ed]
He be Koelman, NHL S enden, Leeuwa den/The Ne he lands, he be .koelman@nhls enden.com
Abs ac
Valida ing ship design da a ac oss sys ems is challenging due o agmen ed in o ma ion om mul iple
sou ces, ile ypes, and o ma s – om 2D d awings, 3D models, and speci ica ions, o en ound in
uns uc u ed ex iles. While uni ied 3D models aim o se e as a single sou ce o u h, ensu ing
accu acy and consis ency ac oss all ep esen a ions emains a complex ask. This pape p esen s a
e ie al-augmen ed gene a ion (RAG) solu ion o ex ac ing and compa ing design pa ame e s om
di e se iles and o ma s. The app oach aims o de ec inconsis encies be ween documen s and e sions,
helping designe s main ain da a in eg i y and educe manual e o h oughou he ship design p ocess.
1. In oduc ion and Cos o E o s in Ea ly Ship Design
The ea ly concep design s age in na al a chi ec u e ep esen s a c i ical phase whe e ex ensi e da a
gene a ion occu s unde se e e ime cons ain s. This s age is cha ac e ized by in ensi e mul idiscipli-
na y collabo a ion and compe i i e bidding p ocesses ha equi e simul aneous de elopmen o nume -
ous design documen s ac oss specialized domains, And ews (2018). Despi e he inhe en unce ain y
and eliance on p elimina y es ima ions, concep designs mus apidly con e ge o mee s ingen bid
equi emen s and p ojec imelines. This phase exhibi s unique cons ain s: (1) comp essed ime ames
wi h in ense p essu e, highly collabo a i e wo k lows equi ing specialized expe ise, Le Poole e al.
(2023), con inuous alida ion o design pa ame e s agains pe o mance h esholds, B a haug e al.
(2008), and subs an ial unce ain y in design assump ions and calcula ions, Jo ge e al. (2018).
The con e gence o hese ac o s c ea es an en i onmen highly suscep ible o e o s ha can p opaga e
h ough subsequen design phases, wi h co ec ion cos s escala ing exponen ially as de ail le els in-
c ease, as shown in Fig.1, DeNucci and Hopman (2021). Resea ch indica es ha ea ly-s age design
e o s can esul in cos o e uns when disco e ed du ing de ailed design o cons uc ion phases, and
some imes i epa able e o s leading o high epe cussions, And ews (2021), Rig e ink (2014). This
sensi i i y necessi a es obus alida ion mechanisms o ensu e pa ame e consis ency and accu acy
h oughou he i e a i e design p ocess.
Cu en indus y p ac ice elies hea ily on adi ional design spi al me hodologies and concep a ia ion
me hods (CVM) ha in ol e mul iple manual e iew cycles and e sion synch oniza ion p ocesses,
Papanikolaou (2018). Howe e , hese app oaches ace undamen al limi a ions in mode n ship design
en i onmen s, cha ac e ized by he agmen ed na u e o design in o ma ion ha spans 2D d awings,
3D models, speci ica ions, and uns uc u ed ex iles, B onson e al. (2024). While uni ied 3D and
collabo a i e en i onmen s a e inc easingly p omo ed as single sou ces o u h, ensu ing accu acy and
consis ency ac oss all design ep esen a ions emains a complex challenge, Koelman e al. (2024). S ud-
ies e eal ha enginee s spend app oxima ely 14% o hei ime loca ing in o ma ion and e i ying
accu acy, ep esen ing a signi ican ine iciency in ime-c i ical design phases, Chui e al. (2023). Ex-
is ing e sion con ol and change acking mechanisms p o e inadequa e o managing he apid i e a-
ion cycles cha ac e is ic o concep design. Mo eo e , cu en alida ion app oaches equi e ex ensi e
manual synch oniza ion be ween di e en sys ems and ile o ma s.
172
Fig.1: Rela ionship o commi ed cos s and design eedom (adop ed om Ma is and DeLau en is
(2000))
This pape p esen s a no el app oach in ol ing e ie al-augmen ed gene a ion (RAG) ha add esses
hese undamen al challenges by au oma ically ex ac ing and compa ing design pa ame e s om di-
e se ile ypes and o ma s o e eal da a inaccu acies and po en ial e o s. Unlike adi ional synch o-
niza ion-based solu ions ha equi e comp ehensi e wo k low es uc u ing and leng hy implemen a-
ion pe iods, his me hod ocuses on e iewing inconsis encies ac oss documen s and e sions while
p ese ing exis ing design p ocesses. The key ad an age o his app oach is i s non-dis up i e in eg a-
ion wi h he cu en wo k low; designe s can con inue using hei p e e ed ools and es ablished p ac-
ices, while he solu ion is a ailable o designe s o alida e pa ame e consis ency.
2. Cu en P ac ice and Challenges
Despi e he p oli e a ion o ad anced digi al design en i onmen s, ea ly-s age ship design alida ion
emains hea ily elian on manual c oss- e e encing o he e ogeneous da a sou ces. Designe s a e o en
equi ed o consul and compa e design pa ame e s om CAD models and hyd os a ic calcula ions o
sp eadshee s and egula o y documen s. This p ocess is no only labo -in ensi e and e o -p one bu
also cons ained by ool in e ope abili y and limi ed access o p op ie a y so wa e pla o ms.
A co e challenge lies in he agmen a ion o design da a ac oss mul iple o ma s and sys ems. C i ical
pa ame e s - including p incipal dimensions, o m coe icien s, s abili y ma gins (e.g., GM), and design
coe icien - a e dis ibu ed ac oss lines plans, gene al a angemen s (GA), s uc u al models, weigh
es ima es, and machine y speci ica ions. These pa ame e s exhibi s ong in e dependencies: o
example, hull o m cha ac e is ics in luence hyd os a ic s abili y; s uc u al a angemen s a ec weigh
dis ibu ion; p opulsion equi emen s impac hull esis ance and uel consump ion. Ensu ing cohe ence
ac oss hese dimensions necessi a es con inuous c oss-documen alida ion, which cu en wo k lows
do no adequa ely suppo .
Mo eo e , egula o y compliance u he complica es alida ion. Requi emen s elucida ion is a co e
ask ha in ol es he syn hesis o mul iple egula ions and guidelines om class. Design p oposals
mus align wi h di e se and e ol ing s anda ds, including SOLAS, MARPOL en i onmen al
egula ions, Ene gy E iciency Design Index (EEDI) h esholds, and classi ica ion socie y ules. These
o e lapping equi emen s gene a e a mul i-objec i e alida ion landscape in which inconsis encies can
p opaga e unno iced, pa icula ly when alida ion elies on manual inspec ion.
173
While mode n so wa e such as CADMATIC and AVEVA Ma ine suppo s model-based app o als,
AVEVA (2020), Yllikäinen (2019), hey a e p ima ily op imized o de ailed design and app o al s ages
- no ea ly-s age concep design. Mos ools alida e geome y and compliance bu o e look consis ency
ac oss unc ional pa ame e s. These and o he alida ion echniques a e discussed below:
1. Manual o in-house Valida ion Tools (Isola ed) - Designe s mus manually ex ac and compa e
pa ame e s om echnical documen s (e.g., line plans, hyd os a ic epo s, sp eadshee s). This
ask is no only ime-consuming bu also highly suscep ible o human e o , especially as design
i e a ions inc ease. Isola ed sc ip s o digi al checklis s may help au oma e he alida ion o
speci ic pa ame e s (e.g., GM, LCG/LCB). While help ul, hese solu ions may no scale due o
in e ope abili y limi a ions.
2. Quali y Assu ance (QA) P ocedu es (Pee /Ex e nal Re iew) - In many i ms, alida ion is de-
e ed o QA e iews. These e iews equi e c oss- unc ional eams o manually syn hesize
inpu s ac oss disciplines, inc easing he cogni i e load. Al hough he e is new esea ch in his
domain, he e is a need o c i ical company buy-in o hese QA p ocesses and equi e dedi-
ca ed pe sonnel o ca y h ough, Hmeshah e al. (2015).
3. Model-Based Valida ion (OCX and Simila S anda ds) - OCX-based wo k lows and 3D model-
cen ic pla o ms a e designed o encapsula e alida ion wi hin a uni ied geome ic model.
Howe e , hese models ypically suppo only hose pa ame e s ha can be di ec ly isualized
o geome ically mapped (e.g., s uc u al membe s, a angemen bounda ies). Alphanume ic
pa ame e s such as s abili y ma gins, pe o mance coe icien s, o o he impo an design da a
cu en ly s ill emains ou side he scope o hese models and mus be alida ed sepa a ely,
As up (2022).
4. So wa e Tools – Class is also leading he de elopmen o new ools o alida ion. Fo exam-
ple, he de elopmen o Nau icus Hull’s Rules Check allows use s o un hei ini e elemen
analysis (FEA) agains ele an ca go holding ules and h esholds, DNV GL (2018). These
ools, apa om DNV Nau icus Hull, include AMBS Eagle UDM, ClassNK P imeShip-Hull,
Lloyd's Regis e ’s RulesCalc, Ko ean Regis e ’s SeaT us -HullScan, among o he s. Howe e ,
hese a e mainly ocused on s uc u al alida ion.
5. No el app oaches – New solu ions a e being p oposed by pe sons such as Soman (2015), who
aim o imp o e he Sma Ship Design (S3D) en i onmen by add essing he cu en lack o
capabili y in e alua ing design agains enginee ing guidelines. The p oposed solu ion uses Na -
u al Language P ocessing (NLP) o ex ac design guidelines e icien ly. Howe e , he solu ion
s ops a he ex ac ion le el, Soman e al. (2015), a gap his pape hopes o add ess.
3. La ge Language Models (LLMs) and RAG
La ge Language Models (LLMs), when combined wi h RAG, open new possibili ies o assessing
inconsis encies ac oss echnical documen s. While LLMs p o ide con ex -awa e easoning o e
complex language, RAG enhances his capabili y by inco po a ing esh, ex e nal da a in o he model’s
esponses. By embedding and indexing echnical documen s, he sys em can ins an ly c oss- e e ence
hem - allowing ship designe s o ask ques ions such as, “Is he GM alue consis en ac oss all epo s?”
o “Does his hull design mee SOLAS and EEDI s anda ds?”
La ge p e ained language models a e highly e ec i e a e aining knowledge and e ie ing ac ual
in o ma ion om hei pa ame e s. Howe e , hei e ec i eness ends o dec ease on downs eam asks
ha equi e expanding o upda ing hei knowledge. Hyb id app oaches ha combine pa ame ic
memo y wi h non-pa ame ic memo ies can help add ess hese limi a ions, as hey allow knowledge o
be e ised and expanded mo e easily and quickly, Lewis e al. (2020). Siddha h and Luo (2024)
in oduce a e ie al-augmen ed gene a ion (RAG) amewo k speci ically designed o ship design
pa en s. I ocuses on ex ac ing named en i ies and hei ela ional s uc u es om pa en ex s o
cons uc a s uc u ed, domain-speci ic knowledge base ha suppo s mo e accu a e and con ex - ich
in o ma ion e ie al.
174
Fig.2 illus a es he enhancemen suppo ed by RAG in he 3-s ep p ocess o p omp o ques ion
answe ing, co e ing indexing o documen s, e ie al o ele an documen s based on seman ic
simila i y, and inpu in o he LLM o he gene a ion o inal answe s. Exis ing esea ch, such as Soman
(2015), is limi ed o il e ing, highligh ing, and ex ac ing ele an ules om echnical s anda ds. By
inco po a ing he gene a i e easoning o LLMs oge he wi h RAG’s abili y o e ie e and in eg a e
ex e nal knowledge, hese capabili ies can be signi ican ly ex ended. This enables ship designe s no
only o iden i y applicable ules bu also o c oss- alida e hem agains new documen s and e ol ing
designs.
Fig.2: Rep esen a ion o he RAG p ocess enhancing LLM o p omp ing, Gao e al. (2023)
4. Case S udy
To explo e he po en ial o LLM-RAG alida ion app oaches in ea ly-s age ship design, we ocus on a
case s udy cen e ed on he esea ch essel ‘RV Gunne us’. This case examines whe he inconsis encies
in design pa ame e s can be e ec i ely iden i ied ac oss mul iple documen ypes and subsequen ly
e iewed agains es ablished design ules.
In his pipeline, we conside he usabili y o LLMs and RAG mainly in:
• Ex ac ing pa ame e s om uns uc u ed PDFs and ex iles
• C oss- alida ing ac oss in e nal design documen s
• Highligh ing inconsis en and egula o y equi emen s,
4.1. Pa ame e s and Da ase
The s udy ocuses on a co e se o in e ela ed pa ame e s commonly ound in speci ica ion shee s such
as p incipal dimensions (e.g., Leng h O e all, Beam, D a ), capaci ies, machine y da a, equipmen ,
and mission-speci ic acili ies. Hence, he da ase used in ol ed he ‘RV Gunne us’ speci ica ion
shee s, gene al a angemen (GA) d awings, hyd os a ics inpu s used o p elimina y hyd odynamics
es s, he 3D model, and equipmen da a. This o iginal da ase con ains: (1) he 3D model, (2) echnical
2D d awings, and (3) ex da a in PDFs. The da ase was p o ided in pa h ough esea ch wi hin he
SEUS P ojec , which enabled access o NTNU ShipLab. Fo his case s udy, he da a was used in i s
o iginal o m wi hou any p e-p ocessing.
175
In he nex sec ion, we discuss he key echnical aspec s o de eloping he AI assis an , including da a
p epa a ion, sys em de elopmen , and model selec ion.
4.2. Me hods
Fig.3: Me hods o 2-s ep Valida ion o Pa ame e s
The me hods o his s udy employ a 2-s ep app oach, as isualized in Fig.3, o co e he a o e-
men ioned goals:
1. Da a Valida ion Tes s be ween Files: Fo his s ep, he con en s o he a ious design iles a e
compa ed wi h each o he . The iles ange om 2D, ex , o 3D d awings and a e p e-p ocessed
wi h he help o an LLM and u he ed back in o he model o p omp es ing. In o de o
unde s and he gains in using he LLM-RAG model, we compa e he esul s o his s ep wi h a
Py hon sc ip ha au oma es he compa ison o pa ame e s be ween iles.
2. Compliance and Valida ion o design pa ame e s agains ules: Fo his s ep, he e ed pa am-
e e s a e hen compa ed o ulese s – an addi ional s anda d p o ocol o compliance and design
alida ion. Gi en ha RV Gunne us has known no a ions, addi ional guidelines a e ed in o he
model and used o assess whe he he pa ame e s a e po en ially complian o no .
4.2.1. Da a P ocessing
Fo da a p ocessing, he goal was o de elop a scalable pipeline sui able o companies and o ganiza-
ions handling la ge olumes o da a. Some o he da a was al eady in ex o ma , bu a signi ican
po ion i s had o be con e ed in o images and hen ex ac ed as ex , since he a ailable me ada a
was no use ul. To add ess his, we employed au oma ed app oaches using Py hon, wi h ools such as
docx, py esse ac , PIL's Image module, and PyMuPDF ( i z).
As pa o he da a p ocessing pipeline, we b ie ly in es iga ed he in eg a ion o 3D model da a in o
ou sys em. The da ase p o ided con ained p ima ily. p and .x_ iles, which a e p op ie a y o ma s
ypically c ea ed wi h licensed so wa e such as Siemens NX. These o ma s could no be p ocessed
di ec ly using open-sou ce Py hon lib a ies like py honOCC. Howe e , we iden i ied con e ing hese
iles in o mo e accessible o ma s such as. STEP (.s p), STL, o . IGES would enable u he p ocessing
176
and analysis. Since Siemens NX suppo s sc ip ing o la ge-scale ba ch con e sions, his con e sion
s ep can be inco po a ed in o an au oma ed wo k low wi hou comp omising he scalabili y o he
pipeline. To ully in eg a e he 3D models in o he RAG amewo k, we need o de ine he ele an
keywo ds o s uc u al ea u es o ex ac om he 3D da a o suppo meaning ul e ie al and
knowledge augmen a ion. Achie ing his would also equi e de eloping addi ional sc ip s o
calcula ing dimensional p ope ies. Howe e , o keep he scope o his pape ocused and o a oid
po en ial e o s om miscalcula ions, we limi ou wo k he e o 2D PDFs and ex da a.
4.2.2. Da a Compa isons
Fo compa ing he da a o he iles, we de eloped a Py hon sc ip o complemen he RAG me hod and
compa e he e ec i eness o he RAG sys em. The Py hon sc ip collec s he p e-p ocessed da a,
appends hem in o a da a ame, and pi o s he da a ame such ha only unique pa ame e s a e indexed
and he di e en sou ces a e conca ena ed in o columns. These alues a e hen compa ed in o de o
de e mine i hey a e consis en o inconsis en . The ou pu esul s e eal a summa y epo in ex ile
o ma .
Fig.4: La es Voyage AI Mul imodal F amewo k, VoyageAI (2024)
4.2.3. RAG Sys em
In designing he RAG sys em, we iden i ied he need o wo dis inc machine lea ning models: one o
embedding he da a in o a ec o space o e icien s o age and e ie al, and ano he as an LLM o
gene a e cohe en ex ual ou pu s. Gi en he mul ilingual na u e o he da ase , i was essen ial o ensu e
ha he sys em e ie es seman ically ele an con en based on con ex a he han language simila i y.
Th ough p elimina y e alua ions, he Voyage Mul ilingual model demons a ed he bes pe o mance
in e ie ing con ex ually accu a e in o ma ion ac oss languages, VoyageAI (2024). Fo ex gene a ion,
we chose GPT-4o due o i s s a e-o - he-a mul ilingual capabili ies, low la ency, and cos -
e ec i eness. These decisions build on he ounda ions laid by he o iginal GPT-4 a chi ec u e, which
demons a ed obus mul ilingual easoning and gene a ion ac oss asks, Open AI e al. (2023).
4.2.4. Web In e ace
The web in e ace was de eloped using Py hon and FAISS o he backend, and Boo s ap and
Ja aSc ip o he on end. The sys em is designed o p o ide a cha -based en i onmen whe e domain
expe s can pose ques ions and ecei e esponses gene a ed om con ex ually ele an documen s. Key
componen s o he sys em a e ully con igu able, allowing use s o selec he numbe o op e ie ed
documen s (n- op), he simila i y me ic, he embedding model, and he language model (LLM) used
o gene a ion. A documen iewe is in eg a ed in o he igh -hand panel, enabling use s o alida e
he gene a ed answe s by e iewing he sou ce documen s. These documen s can also be downloaded
o u he inspec ion.
177
Fig.5: Web In e ace o he RAG Solu ion
4.3. Resul s
Tes s o he i s s ep e eal ha abou 173 unique pa ame e s we e ound. In he Py hon baseline
app oach, abou 168 we e known consis en pa ame e s and abou 5 pa ame e s we e deemed
inconsis en . These pa ame e s include moulded b ead h, ca go hold olume, uel oil olume, onnage,
and dep h da a. The ou pu summa y ile is shown in Fig.6. Howe e , he e a e limi a ions o his
me hod in ha seman ically equi alen pa ame e s we e no compa ed. Fo ins ance, ‘loa’ was no
compa ed o leng h o e all.
In compa ison, he LLM model claimed o de ec abou 16 inaccu acies om he pos -p ocessed 173
unique pa ame e s and was able o e iew seman ic simila i ies, gi en he p omp : ‘wha inconsis encies
do you ind’. Fig.7 shows hese pa ame e s. I caugh he same inconsis encies lagged by he Py hon
sc ip bu also wen u he by iden i ying seman ically simila e ms and compa ing hem. Fo example,
i co ec ly ecognized ha LOA and leng h o e all e e o he same measu e, as wi h LWL and leng h
a wa e line. I highligh ed he di e ences in hese alues in he speci ica ion shee and GA d awing.
This disc epancy is likely due o he GA ep esen ing a di e en (leng hened) e sion o he essel
compa ed o he speci ica ion. The model was also able o dis inguish be ween di e en ypes o wa e
capaci y, such as echnical wa e e sus po able wa e , and compa e hei alues ac oss documen s -
e ealing, o ins ance, a 0.4 m³ di e ence be ween he speci ica ion shee and he GA.
Howe e , he model also p oduced some less meaning ul compa isons. In se e al cases, i compa ed
e ms agains hemsel es (e.g., d augh unde side keel, ne onnage, im, and wa e ballas olume),
which esul ed in misleading ou pu s. I also compa ed d augh no ma i e wi h d augh a max load
(concep ually di e en measu es), and inconsis encies in he le el o de ail when desc ibing he same
p opulsion equipmen .
O e all, we ind he RAG solu ion can de e mine mo e equi alen pa ame e s, bu we also obse e ha
i ends o o e -co ec , showing high sensi i i y o anomalies and o en a emp ing o in e mo e
di e ences han a e ac ually p esen . Compa ed wi h he exis ing pipeline, we expec he RAG unc ion
o s eamline he p ocess by emo ing he need o manual execu ion o a Py hon sc ip , o e ing g ea e
con enience h ough he de eloped in e ace. The use - iendly in e ace also acili a es easie sc u iny
178
o esul s, unlike he Py hon sc ip whe e compa isons a e ha dcoded and equi e use s o e iew he
code di ec ly o ensu e no hing was o e looked.
Fig.5: Resul s om Py hon Sc ip
On op o he ex ac ion piece, as pe he es s o S ep 2, we a e also hoping o es how he RAG can
be used o e alua e compliance o he gi en pa ame e s agains he ules. The essel has he ollowing
ule no a ions: DNV + 1A1 + Ice C + E0 + R2 Ca go ship. The goal was o double-check how much
he RAG can suppo design compliance, gi en he known and alida ed pa ame e s agains he ules.
The model was popula ed wi h DNV ules DNVGL-RU-SHIP Pa s 1 o 6 and he ollowing p omp
was an: ‘Does he RV Gunne us speci ica ion comply wi h he a ached DNV ules o a esea ch
essel, gi en ha i has he ollowing ule no a ions <<DNV + 1A + Ice C + E0 + R2 Ca go ship>>?’
Wi h he aid o b eaking he p omp u he down, he RAG-LLM was able o asce ain i s wha he
no a ions means: ‘The no a ion "1A + Ice C + E0 + R2" indica es ha he ca go ship is classi ied wi h
he ollowing speci ica ions: "1A" signi ies a high ice class o essels ope a ing in ice-in es ed wa e s,
"Ice C" deno es he ship's capabili y o na iga e in ligh ice condi ions, "E0" indica es he essel has no
es ic ions on he use o elec ical p opulsion, and "R2" e e s o he ship's compliance wi h speci ic
equi emen s o eliabili y and edundancy in i s sys ems.’ Howe e , o he speci ica ions o each
no a ion, he model ad ises consul ing he la es DNV classi ica ion o he la es guidance, showing
ha while i can asce ain he ules, i is no able o si h ough he pa ame e s whe e h esholds we e
ob ious. I is impo an o no e ha , a his s age, we do no expec he LLM o ha e he abili y o
pe o m mo e complex ma h and calcula ions in o de o in e compliance.
179
Fig.6: Resul s o STEP 1 and co esponding ulese s in he con ex sideba
Fig.7 displays he in e ace s a ing wi h he p omp s a ound design pa ame e inconsis encies, as
in oduced. The Con ex Menu demons a es how hese p omp s a e connec ed o he ule se s p o ided
o he model. As addi ional ules and design da a a e inco po a ed, he in e ace has he po en ial o
se e as a pla o m whe e designe s can no only de ec da a inconsis encies bu also e alua e hem
agains he co esponding ules in he Con ex Menu – making he wo alida ion s eps no only possible
bu pe o med in pa allel.
4.4. Discussions and nex s eps
The e a e a ious limi a ions and lea nings om he case s udy ha a e subjec o u u e imp o emen s
bo h o he model and he in e ace:
1. Me ada a: Cu en ly, he model only eads ex and con ex da a o e ie al. This makes i
challenging o p omp he LLM o e iew mo e speci ic iles and documen s. The me ada a o
he iles, including ile name, e sion, and da e o gene a ion, is no cu en ly conside ed, bu
would be help ul o u u e e e ence, allowing designe s o easily poin o speci ic documen s.
2. Uploading in e ace: Alongside his de elopmen , a mo e use - iendly in e ace o uploading
documen s and e ie ing hem would be handy. To make es ing and explo a ion easie , i
would be help ul o add a d op box o uploading iles ha au oma ically upda es he da abase.
3. Secu e da abase o co po a e da a: Le e aging co po a e da a and deploying local ins ances o
he model can ailo he LLM o a company’s speci ic needs. By aining i on co po a e em-
pla es, designs, and e minology, he model becomes be e aligned wi h o ganiza ional p ac-
ices, making p omp s o essel ypes, p ojec numbe s, and o he domain-speci ic inpu s mo e
in ui i e and cus omized o he design eam.
No ing hese po en ial imp o emen s, he addi ion o mo e and mo e da a o aining he LLM can
inc ease he o e all sensi i i y and accu acy o he model. Fu he aining is expec ed o he model so
ha i can become inc easingly awa e o ma i ime and ship design-speci ic seman ics.
