{"id":{"repo_id":"vilnius","oai_identifier":"oai:vu.lt:elaba:308336747"},"canonical_url":"https://search.ndltd.org/etd/vilnius/oai:vu.lt:elaba:308336747","repository":{"repo_id":"vilnius","name":"Vilnius University","base_url":"https://epublications.vu.lt/oai"},"display":{"title":"Daugiaagentė sistema, naudojanti didelės kalbos modelius verslo procesams automatizuoti /","abstract":"Invoice processing is an expensive part of accounting: some Lithuanian companies pay up to 0,30 EUR for each invoice to be processed, while others send the documents to „Microsoft Document Intelligence“ or „Google Document AI“, thus letting the data out of company boundaries. This paper analyses a third variant: can a locally&#8209;hosted open&#8209;source model replace these solutions. Experiment uses 33 real invoices to test seven local models (from 2 billion parameter „Gemma 4 E2B“ till 27 billion parameter „Qwen 3.6 27B“) and two cloud models („GPT&#8209;5.4&#8209;nano“, „GPT&#8209;5.5“) in five scenarios, including a multi&#8209;agent system with extractor, judge, formatter agents. The best local model („Mistral Devstral Small 2 2512“) with „Tesseract“ OCR help achieves the same F1 score as „GPT&#8209;5.5“ (&#8776;0,90), while using only half of the output tokens. Smaller models’ accuracy is increased by 0,16&#8209;0,18 adding OCR, reasoning add no additional value, and parallel processing increases throughput up to 60 % without accuracy cost. Total cost of the experiment was &#8776;4.63 EUR, 20 to 540 times cheaper than the same workload would have been on cloud services.","abstract_html":"Invoice processing is an expensive part of accounting: some Lithuanian companies pay up to 0,30 EUR for each invoice to be processed, while others send the documents to „Microsoft Document Intelligence“ or „Google Document AI“, thus letting the data out of company boundaries. This paper analyses a third variant: can a locally&amp;#8209;hosted open&amp;#8209;source model replace these solutions. Experiment uses 33 real invoices to test seven local models (from 2 billion parameter „Gemma 4 E2B“ till 27 billion parameter „Qwen 3.6 27B“) and two cloud models („GPT&amp;#8209;5.4&amp;#8209;nano“, „GPT&amp;#8209;5.5“) in five scenarios, including a multi&amp;#8209;agent system with extractor, judge, formatter agents. The best local model („Mistral Devstral Small 2 2512“) with „Tesseract“ OCR help achieves the same F1 score as „GPT&amp;#8209;5.5“ (&amp;#8776;0,90), while using only half of the output tokens. Smaller models’ accuracy is increased by 0,16&amp;#8209;0,18 adding OCR, reasoning add no additional value, and parallel processing increases throughput up to 60 % without accuracy cost. Total cost of the experiment was &amp;#8776;4.63 EUR, 20 to 540 times cheaper than the same workload would have been on cloud services.","abstract_has_math":false,"creators":["Džiaugys, Nojus,"],"institution":"Institutional Repository of Vilnius University","degree_name":null,"degree_level":null,"degree_discipline":null,"degree_department":null,"school":null,"contributors":[],"advisors":[],"committee_chairs":[],"committee_members":[],"year":2026,"date_issued":"2026","date_published":"2026","updated_at":"2026-11-19T10:12:55Z","subjects":[],"languages":["lit"],"rights":["info:eu-repo/semantics/openAccess"],"rights_urls":[],"identifier_entries":[],"person_identifier_entries":[]},"links":{"outbound_url":"https://repository.vu.lt/VU:ELABAETD308336747&prefLang=en_US","outbound_label":"Repository record","outbound_source":"dc:identifier","record_url":"https://repository.vu.lt/VU:ELABAETD308336747&prefLang=en_US","record_label":"Repository record","record_kind":"record","record_source":"dc:identifier","download_url":null,"download_label":null,"download_source":null,"access_status":"open","embargo_until":null,"access_note":null},"source_record":{"url":"https://epublications.vu.lt/oai?verb=GetRecord&metadataPrefix=oai_dc&identifier=oai%3Avu.lt%3Aelaba%3A308336747","prefix":"oai_dc"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:creator","label":"Author","values":["Džiaugys, Nojus,"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date","label":"Dc Date","values":["2026"]},{"key":"dc:publisher","label":"Institution","values":["Institutional Repository of Vilnius University"]},{"key":"dc:relation","label":"Dc Relation","values":["https://epublications.vu.lt/object/elaba:308336747/308336747.pdf"]},{"key":"dc:type","label":"Dc Type","values":["info:eu-repo/semantics/bachelorThesis"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:language","label":"Dc Language","values":["lit"]},{"key":"dc:rights","label":"Dc Rights","values":["info:eu-repo/semantics/openAccess"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier","label":"Identifier","values":["https://repository.vu.lt/VU:ELABAETD308336747&prefLang=en_US"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description","label":"Description","values":["Invoice processing is an expensive part of accounting: some Lithuanian companies pay up to 0,30 EUR for each invoice to be processed, while others send the documents to „Microsoft Document Intelligence“ or „Google Document AI“, thus letting the data out of company boundaries. This paper analyses a third variant: can a locally&#8209;hosted open&#8209;source model replace these solutions. Experiment uses 33 real invoices to test seven local models (from 2 billion parameter „Gemma 4 E2B“ till 27 billion parameter „Qwen 3.6 27B“) and two cloud models („GPT&#8209;5.4&#8209;nano“, „GPT&#8209;5.5“) in five scenarios, including a multi&#8209;agent system with extractor, judge, formatter agents. The best local model („Mistral Devstral Small 2 2512“) with „Tesseract“ OCR help achieves the same F1 score as „GPT&#8209;5.5“ (&#8776;0,90), while using only half of the output tokens. Smaller models’ accuracy is increased by 0,16&#8209;0,18 adding OCR, reasoning add no additional value, and parallel processing increases throughput up to 60 % without accuracy cost. Total cost of the experiment was &#8776;4.63 EUR, 20 to 540 times cheaper than the same workload would have been on cloud services."]},{"key":"dc:format","label":"Dc Format","values":["application/pdf"]},{"key":"dc:title","label":"Title","values":["Daugiaagentė sistema, naudojanti didelės kalbos modelius verslo procesams automatizuoti /","Multi-agent system using large language models for business process automation."]}]}],"canonical_facts":{"dc:creator":["Džiaugys, Nojus,"],"dc:date":["2026"],"dc:description":["Invoice processing is an expensive part of accounting: some Lithuanian companies pay up to 0,30 EUR for each invoice to be processed, while others send the documents to „Microsoft Document Intelligence“ or „Google Document AI“, thus letting the data out of company boundaries. This paper analyses a third variant: can a locally&#8209;hosted open&#8209;source model replace these solutions. Experiment uses 33 real invoices to test seven local models (from 2 billion parameter „Gemma 4 E2B“ till 27 billion parameter „Qwen 3.6 27B“) and two cloud models („GPT&#8209;5.4&#8209;nano“, „GPT&#8209;5.5“) in five scenarios, including a multi&#8209;agent system with extractor, judge, formatter agents. The best local model („Mistral Devstral Small 2 2512“) with „Tesseract“ OCR help achieves the same F1 score as „GPT&#8209;5.5“ (&#8776;0,90), while using only half of the output tokens. Smaller models’ accuracy is increased by 0,16&#8209;0,18 adding OCR, reasoning add no additional value, and parallel processing increases throughput up to 60 % without accuracy cost. Total cost of the experiment was &#8776;4.63 EUR, 20 to 540 times cheaper than the same workload would have been on cloud services."],"dc:format":["application/pdf"],"dc:identifier":["https://repository.vu.lt/VU:ELABAETD308336747&prefLang=en_US"],"dc:language":["lit"],"dc:publisher":["Institutional Repository of Vilnius University"],"dc:relation":["https://epublications.vu.lt/object/elaba:308336747/308336747.pdf"],"dc:rights":["info:eu-repo/semantics/openAccess"],"dc:title":["Daugiaagentė sistema, naudojanti didelės kalbos modelius verslo procesams automatizuoti /","Multi-agent system using large language models for business process automation."],"dc:type":["info:eu-repo/semantics/bachelorThesis"]},"updated_at":"2026-11-19T10:12:55Z"}