Trallie: Shaping Unstructured Data Into Valuable Information
summary
Open-source framework Trallie converts file collections (PDF, HTML, TXT, JSON) into validated JSON using LLMs (GPT‑4o, Grock, Llama, Mistral, Hugging Face endpoints) and ships as a pip package. It auto-generates a schema from a small sample (default 10 docs) and then extracts values at scale, with JSON resampling up to max_retries=5. Benchmarks on a few‑hundred‑document FDA set show closed IE extraction is reliable even zero/few‑shot (70B+ models best) while schema generation remains the harder, error‑prone step.
AI-written summary · about
chapters (25)
- 0:00Trallie and Pi School intro
- 1:01Unstructured data challenges
- 2:27Trallie framework overview
- 3:31Schema generator explained
- 6:24Schema generation harder than extraction
- 7:28Data extractor walkthrough
- 8:12Demo setup and pip install
- 9:37Trallie with local files
- 10:41Starter schemas and examples
- 11:46Schemas from raw text
- 12:28Extracting with predefined schemas
- 13:11JSON parsing and hallucinations
- 13:51Image formats and OCR limits
- 15:40Evaluating extraction systems
- 16:47Token vs field-based evaluation
- 18:36Supported models and multilingual
- 19:40FDA benchmark results
- 20:44Open IE challenges
- 21:28Future benchmarks and evaluations
- 22:12Q&A
- 22:34Multilingual and sensitive data
- 24:51Privacy for local models
- 25:36Flat schemas and graph support
- 26:45PDF parsing limits
- 27:05Closing remarks
related talks
01h 00mComplex Data Ingestion with Open Source AI
01h 25mLearn to Unlock Document Intelligence with Open-Source AI
01h 19mLearn to Unlock Document Intelligence with Open-Source AI
26mAdvanced Document Parsing for AI Workflows
29mConquering PDFs- document understanding beyond plain text
23mSynthetic data that meets enterprise compliance needs
29mSchemas: The Key To Data Happiness
34mBuilding Trust with LLMs