← Back to home

Tiny LLM Record Processor

Paste one record per line using date|group|body|to|from. A small local model classifies each record, and a filtered RAG-style summary runs with a Hugging Face model downloaded into your browser.

Allowed users in to and from: stu, bob, sally.

Input Records

Each non-empty line must contain 5 fields: YYYY-MM-DD|group|body|to|from.

Record lines User filter (matches to OR from)

Hugging Face browser model (downloaded for summary) First run downloads model files to browser cache, then reuses them.

Entity extraction model Choose local rules or browser NER model for entity extraction.

Model Output

Local model not loaded yet.

Remote model: Not loaded.

Entity model: Local regex extractor.

No summary generated yet.

#	Date	Group	To	From	Intent	Confidence	Suggested To	Entities	Body
No records processed yet.

Filtered summary

Run “Summarize filtered (RAG style)” to generate a summary.

Extracted entities (filtered set)

Run “Process records” to extract entities.

How this page works: Records are parsed as date|group|body|to|from, validated, and scored by a lightweight local model for intent and suggested routing. Filtering includes rows where the selected user appears in to or from. The page extracts entities from each body with either local regex rules or a selected browser NER model (including dslim/distilbert-NER) and aggregates them for the filtered set. For summarization, it retrieves filtered body text into a compact context window (RAG retrieval step), then runs a Hugging Face model downloaded in-browser via transformers.js. If model load or inference fails, local fallback logic is used.