LM-Kit OneDocs2026.8.10lm-kit.com
Documents

Text Analysis

Ten endpoints under /lmkit/v1 answer questions ABOUT content: what it is, what it names, what it reveals, how it reads. Two facts make them more useful than their one-line summaries suggest: nearly all of them accept whole DOCUMENTS (PDF, HTML, EML, MBOX, Office, images with OCR), not just strings, and the classification and entity endpoints are shaped by YOUR labels at request time, so "custom NLP" here means writing a request, not training a model.


1The toolbox#

Endpoint Answers The lever that makes it yours
categorize Which of YOUR categories fits Your taxonomy at request time; the full treatment is Document Classification
ner Which entities are named built_in_types selection plus custom_labels for domain entities
pii-detection What personal data is present Same shape: built-in types plus custom labels
sentiment-analysis Tone: sentiment, emotion, sarcasm Each dimension toggles independently, with neutral support
keyword-extraction The terms that carry the content Feeds tagging and lightweight retrieval
language-detection What language this is The router in front of translation and language-specific handling
translation The content in another language Whole documents, structure preserved
summarization The short version Length budgets, optional title generation, intent, and target_language for cross-lingual summaries
text-correction Grammar and spelling fixed Correction without rewriting
text-rewriter The same content, said differently Tone and style transformation

Exact request shapes live in the API reference; models resolve like everywhere else (the request's model, or the default slot for the capability).

2The two request patterns worth knowing#

Documents in, not strings. An invoice PDF, an email thread, or a screenshot goes into categorization or pii-detection directly: the document pipeline (extraction plus OCR) runs inside the endpoint. You only pre-convert when you want to reuse the extracted text across several calls, in which case upload once and pass the file id everywhere.

Labels at request time. NER and PII detection take custom_labels beside the built-in types: contract clause types, product codes, domain-specific identifiers, declared per request, plus guidance for interpretation rules. Classification follows the same philosophy with described categories and an honest open-set switch; as the router of the IDP pipeline it has its own chapter.

3Pipelines these endpoints form#

  • Inbox triage: language-detection routes, classification files, ner fills the ticket fields, summarization writes the handler's brief. Four calls per message, all local.
  • Compliance sweep: pii-detection over a document folder, human review of the findings, then true redaction with the confirmed strings. Detection returns what and where; redaction destroys it.
  • Cross-language corpus: language-detection, then summarization with target_language, so a mixed-language archive reads in one language without a separate translation pass.
  • Search enrichment: classification and keyword-extraction at ingest produce the metadata that search filters and facets run on. Decide metadata before bulk-indexing; retrofitting it means re-indexing.

4Choosing models for analysis#

Analysis tasks are the natural home of SMALL models: classification against described categories, entity tagging, and language detection are constrained tasks where a compact model is fast and reliably on-format, and throughput usually matters because these endpoints run per message or per document at volume. Start small, measure against a labeled sample of your own traffic, and escalate only where the small model measurably confuses your actual categories. For a category system with real volume, the escalation path is fine-tuning a compact model on your corrected outputs rather than renting a bigger generic one.

5Stated plainly#

  • Every analysis endpoint reads whole documents through the same pipeline as everything else; OCR and format handling are not your problem.
  • Custom taxonomies are request parameters, and the label DESCRIPTIONS are where the accuracy lives.
  • These endpoints compose: triage, compliance-and-redact, cross-language reading, and search enrichment are each a few calls on one local server.