--- schema_version: "newruntime-agent-readable-v0.2" type: "tectonic_shift" stable_id: "tectonic_shift:document-processing-shifts-from-hand-crafted-ocr-pipelines-to-llm" slug: "document-processing-shifts-from-hand-crafted-ocr-pipelines-to-llm" title: "Document processing shifts from hand‑crafted OCR pipelines to LLM‑driven conversational extraction" description: "Developers are moving away from manually built OCR and schema definitions toward AI‑powered extraction where large language models generate and refine extraction schemas and handle OCR errors in dialogue." retrieval_nugget: "Developers are moving away from manually built OCR and schema definitions toward AI‑powered extraction where large language models generate and refine extraction schemas and handle OCR errors in dialogue. Before: Manual OCR pipelines with hand‑written regex/JSON schemas and static error handling After: LLM‑guided OCR and schema creation via conversational interfaces, with dynamic error correction Current stage: emerging; confidence is medium." status: "published" before_state: "Manual OCR pipelines with hand‑written regex/JSON schemas and static error handling" after_state: "LLM‑guided OCR and schema creation via conversational interfaces, with dynamic error correction" stage: "emerging" confidence: "medium" magic_decay_stage: "useful-workflow" first_seen: "2026-09-26" last_verified: "2026-09-28" record_date: "2026-09-28" date_kind: "last_verified" related_posts: [] topics: [] source_urls: ["https://newruntime.com/signals/llm-ocr-why-the-errors-got-harder-to-e9bd75/","https://newruntime.com/signals/parse-extract-classify-now-each-with-its-own-ce8423/","https://newruntime.com/signals/document-ocr-is-not-getting-commoditized-402503/"] supporting_evidence: [{"title":"LLM OCR: Why the Errors Got Harder to Spot","url":"https://newruntime.com/signals/llm-ocr-why-the-errors-got-harder-to-e9bd75/","kind":"signal","observed_at":"2026-09-27"},{"title":"Parse, Extract, Classify — now each with its own MCP","url":"https://newruntime.com/signals/parse-extract-classify-now-each-with-its-own-ce8423/","kind":"signal","observed_at":"2026-09-26"},{"title":"Document OCR is Not Getting Commoditized","url":"https://newruntime.com/signals/document-ocr-is-not-getting-commoditized-402503/","kind":"signal","observed_at":"2026-09-26"}] counter_evidence: ["In highly regulated industries, organizations may still require deterministic, auditable OCR pipelines that hand‑crafted rules provide, limiting the applicability of LLM‑driven approaches.","Some enterprises report that LLM‑based extraction introduces latency and cost overhead compared to optimized traditional OCR engines for large‑scale batch processing."] routes: {"html":"https://newruntime.com/shifts/document-processing-shifts-from-hand-crafted-ocr-pipelines-to-llm/","markdown":"https://newruntime.com/shifts/document-processing-shifts-from-hand-crafted-ocr-pipelines-to-llm.md","json":"https://newruntime.com/shifts/document-processing-shifts-from-hand-crafted-ocr-pipelines-to-llm.json"} source_format: "knowledge-ledger" published_at: "2026-09-28T06:03:50.831Z" auto_created: true editorial_provenance: {"desk":"knowledge-revision","provider":"groq","model":"openai/gpt-oss-120b","created_at":"2026-09-28T06:03:50.831Z","week_start":"2026-09-21","review":"auto-created from reviewed evidence; the owner may revise or retire it"} --- # Document processing shifts from hand‑crafted OCR pipelines to LLM‑driven conversational extraction ## Retrieval answer Developers are moving away from manually built OCR and schema definitions toward AI‑powered extraction where large language models generate and refine extraction schemas and handle OCR errors in dialogue. Before: Manual OCR pipelines with hand‑written regex/JSON schemas and static error handling After: LLM‑guided OCR and schema creation via conversational interfaces, with dynamic error correction Current stage: emerging; confidence is medium. ## Before and after - Before: Manual OCR pipelines with hand‑written regex/JSON schemas and static error handling - After: LLM‑guided OCR and schema creation via conversational interfaces, with dynamic error correction Stage: emerging. Confidence: medium. Magic decay: useful-workflow. ## What is shifting? **Before:** Manual OCR pipelines with hand‑written regex/JSON schemas and static error handling **After:** LLM‑guided OCR and schema creation via conversational interfaces, with dynamic error correction ## What the evidence shows The three signals describe a consistent shift in document‑processing workflows. LlamaIndex’s note on LLM OCR (e14) highlights how error patterns are changing, implying that traditional rule‑based pipelines are becoming less effective. The Conversational Extract feature (e16) lets users build extraction schemas through dialogue instead of hand‑coding JSON, directly replacing manual schema authoring. Finally, the argument that OCR will not become a generic commodity (e20) underscores the move toward specialized, model‑driven extraction rather than static OCR tools. Together, these items show a pattern where AI models take over both the perception (OCR) and the structuring (schema) steps. This shift is driven by rapid improvements in vision‑language models that can understand document layouts and generate structured outputs on the fly, reducing the need for domain experts to write and maintain brittle extraction rules. The conversational interface lowers the barrier for non‑technical users to define what data they need, while the LLM’s ability to reason about errors makes the pipeline more resilient. Builders should start integrating LLM‑backed extraction APIs into their document pipelines, expose conversational schema editors to end‑users, and design monitoring that captures LLM‑specific error modes. Investing in prompt engineering and feedback loops will help maintain extraction quality as models evolve. ## Where this may not hold - In highly regulated industries, organizations may still require deterministic, auditable OCR pipelines that hand‑crafted rules provide, limiting the applicability of LLM‑driven approaches. - Some enterprises report that LLM‑based extraction introduces latency and cost overhead compared to optimized traditional OCR engines for large‑scale batch processing. ## Revision trigger If multiple large enterprises publicly adopt static OCR and schema pipelines for new projects, showing a sustained preference for rule‑based methods over LLM‑driven extraction, we would reconsider the claim. Published automatically by the New Runtime weekly knowledge desk from the week's evidence and revised every week. ## Supporting evidence - [LLM OCR: Why the Errors Got Harder to Spot](https://newruntime.com/signals/llm-ocr-why-the-errors-got-harder-to-e9bd75/) (signal, 2026-09-27) - [Parse, Extract, Classify — now each with its own MCP](https://newruntime.com/signals/parse-extract-classify-now-each-with-its-own-ce8423/) (signal, 2026-09-26) - [Document OCR is Not Getting Commoditized](https://newruntime.com/signals/document-ocr-is-not-getting-commoditized-402503/) (signal, 2026-09-26) ## Counter-evidence - In highly regulated industries, organizations may still require deterministic, auditable OCR pipelines that hand‑crafted rules provide, limiting the applicability of LLM‑driven approaches. - Some enterprises report that LLM‑based extraction introduces latency and cost overhead compared to optimized traditional OCR engines for large‑scale batch processing. ## Provenance Auto-created by the New Runtime weekly knowledge desk on 2026-09-28 from evidence reviewed in the week of 2026-09-21. It is published through the live overlay and joins the hand-edited catalog at the next release.