@typedef.ai.bsky.social
We are here to eat bamba and revolutionize the world of query engines. The Spark is gone, let's rethink data processing with a pinch of AI
Note: auto-routing is being explored; today you keep full control.
check the repo for more: github.com/typedef-ai/f...
Mix providers (OpenAI, Anthropic) with simple aliases
Use defaults for simple ops; override model_alias for complex ones
Balance cost/latency/quality without extra orchestration
Teams often wire a single model and pay in either cost or quality.
With Fenic, you register multiple models once and select them per call.
fenic's Multiple Model Configuration & Selection lets you pick the right model for each step, cheap where you can, powerful where you must.
Think of it as a per-operator model dial across your pipeline.
Thanks to @danielvanstrien.bsky.social and @lhoestq.hf.co for the collaboration and feedback that made this possible and to David Youngworth you built and maintains the integration!
21.10.2025 18:57 — 👍 0 🔁 0 💬 0 📌 0
A few things you can do with this new integration.
1. Rehydrate the same agent context anywhere (local → prod)
2. Versioned, auditable datasets for experiments & benchmarks
Fenic ❤️ Hugging Face Datasets!
You can now turn any fenic snapshot into a shareable, versioned dataset on @hf.co perfect for reproducible agent contexts and data sandboxes.
Docs: huggingface.co/docs/hub/dat...
"AI confidence is high — but production results still lag."
Our cofounder, Yoni Michael, shares why in CIO.
Read it here 👉 www.cio.com/article/4069...
#CIO #AIinEnterprise #Typedef
Common patterns: multi-step enrichment, RAG prep, nightly jobs with partial recomputes.
for more check the Github repo: github.com/typedef-ai/f...
With fenic, it’s explicit and simple: call .cache() where it matters.
Protect pricey semantic ops (classify/extract) from re-execution
Reuse cached results across multiple downstream analyses
Recover from mid-pipeline failures without starting over
Think of it as checkpointing for LLM workloads: cache after costly ops, restart from there if something fails.
Without caching, teams re-pay tokens and time on retries: flaky APIs, disk hiccups, long recomputes.
fenic's Local Data Caching & Persistence keeps expensive AI steps from rerunning and your pipelines resilient.
24.09.2025 01:38 — 👍 0 🔁 0 💬 1 📌 0Mix providers (OpenAI, Anthropic) with simple aliases
Use defaults for simple ops; override model_alias for complex ones
Balance cost/latency/quality without extra orchestration
Teams often wire a single model and pay in either cost or quality.
With Fenic, you register multiple models once and select them per call.
fenic's Multiple Model Configuration & Selection lets you pick the right model for each step, cheap where you can, powerful where you must.
Think of it as a per-operator model dial across your pipeline.
Why do most AI projects stall?
Because going from prototype → production is HARD.
On Data Exchange, we share how Typedef makes inference-first pipelines actually work at scale.
👉 thedataexchange.media/typedef-fenic/
We’re honored to be featured in AI World Today! 🚀
Our co-founder Yoni Michael shares how Typedef is closing the gap between AI prototypes and production, making inference a first-class data operation.
👉 Read the full interview: www.aiworldtoday.net/p/interview-...
We’re building the AI-native, inference-first infrastructure that powers scalable, production-ready LLM pipelines—no infrastructure headaches, just reliable results. Read more in AIM about how we’re overcoming pilot paralysis: aimmediahouse.com/ai-startups/...
20.09.2025 20:42 — 👍 0 🔁 0 💬 0 📌 0
Fenic brings the reliability of DataFrame pipelines to AI workloads—semantic joins, markdown parsing, transcripts, and more—now strengthened with the 0.3.0 update. Dive into the latest improvements. → www.techzine.eu/blogs/data-m...
20.09.2025 15:57 — 👍 0 🔁 0 💬 0 📌 0
AI fatigue is everywhere. But it’s not inevitable.
In AI Journal, Typedef co-founder Yoni Michael shares how teams can escape “pilot paralysis” and move AI from prototype to production with confidence.
👉 Read the article: aijourn.com/ai-fatigue-i...
Common patterns: review mining, invoice parsing, lead enrichment, spec extraction.
for more, check the GitHub repo: github.com/typedef-ai/f...
Define a Pydantic schema; get type-checked structs (ints, bools, lists, Optionals)
Auto-prompting via function calling / structured outputs (OpenAI, Anthropic)
Use unnest() and explode() to work with the data—no manual JSON wrangling
Most teams hand-roll JSON parsing, brittle regex, and post-hoc validators. That’s slow and error-prone.
With fenic, you keep it declarative.
fenic's Structured Output Extraction turns LLM text into validated tables, directly in your DataFrame.
Think of it as schema-first parsing: you define a Pydantic model; Fenic enforces it and returns structured columns.
Common patterns: doc mining, content ingestion, RAG prep, taxonomy extraction.
for more, including examples and documentation, check: github.com/typedef-ai/f...
Type safety: Embedding/Markdown/JSON columns prevent incompatible ops
Built-ins that matter: normalize, similarity, jq queries for JSON and many more
Less glue: query structured + unstructured togethe,; mix dataframes + SQL + AI in one plan
Most teams treat AI artifacts as loose strings/arrays: schema drift, brittle casting, ad-hoc JSON parsing, and inconsistent similarity math.
In fenic, these are first-class.
fenic's First-Class AI Data Types make embeddings, markdown, and JSON real, typed columns, with the right operations built in.
Think of it as strong types for meaning and structure: safer pipelines, richer queries.
for examples and more information, check: github.com/typedef-ai/f...
16.09.2025 01:38 — 👍 0 🔁 0 💬 0 📌 0Most teams fight drift: regex stacks, ad-hoc prompts, inconsistent tags.
With fenic you can:
Define classes once; get schema-clean, consistent labels
Zero-shot or few-shot with real examples (not just descriptions)
Batching, caching, retries built-in all testable in the same plan.