On 2026‑10‑21 I shared a LinkedIn post describing how I combined Jev with Beacon (by Asymptote Labs) to build a reusable cross‑agent memory layer for engineering lessons extracted from agent runs.
Jev makes it dramatically cheaper to identify which corrections and trajectories from an agent run are reusable across agents.
Beacon by Asymptote Labs continuously captures agent sessions across harnesses and stores their traces in a shared index.
In this workflow, one agent never receives another agent's raw transcript; it receives only the specific engineering lesson produced from the evaluated trace.
The polynomial kernel of degree 2 expands to a dot product in a 6-dimensional space without explicitly visiting that space.
The RBF (Gaussian) kernel computes the dot product in an infinite-dimensional space without explicitly visiting that space.
Marker V2 supports more than 90 languages.
I enable the use_llm option.
On 2026‑10‑08 I announced Datalab Marker v2, an open‑source document parsing pipeline that can turn any PDF, image, doc or PPT into clean Markdown, JSON, or HTML. It targets production workloads with millions of pages, mixed formats, inconsistent layouts, low‑quality scans, and multilingual documents.
- Supported input types: PDF, image, doc, PPT
- Output formats: Markdown, JSON, HTML
Marker v2 runs at up to 23.7 pages per second on a single B200 GPU and supports more than 90 languages.
The GitHub repository is 100% open‑source with 40k stars.
I chose a single‑GPU pipeline to prove that high throughput is achievable without large clusters, lowering cost and complexity for many teams.
Flat index compares the query with every stored vector, selects the K smallest distances, and returns them.
IVF divides the vector space into clusters. Each stored vector is assigned to its nearest cluster centroid.
A probe parameter controls how many clusters it searches, and increasing it usually improves recall but also increases latency.
HNSW stores vectors in a multilayer graph.
The upper layers contain fewer nodes and allow large jumps through the graph.
HNSW often provides low latency and high recall, but the graph requires extra memory.
IVF-PQ combines clustering with product quantization.
Product quantization splits each vector into smaller parts and stores compact codes for them.
ScaNN partitions the dataset and stores quantized vector representations.
pg-jev is an open-source Postgres extension that exposes Jev through SQL functions.
jev() works inside WHERE
jev_prob() returns a probability
jev_choice() selects a label
jev_score() ranks rows across ordered levels.
It needs no vector column or embedding index.
In early October 2026 I posted two X threads (2026‑10‑03 10:22 UTC and 2026‑10‑03 21:58 UTC) showcasing an "insane" use‑case for Jev via the new open‑source Postgres extension pg‑jev. The posts walk through installing the extension, using the four SQL functions (jev(), jev_prob(), jev_choice(), jev_score()) to run semantic filters, classifications, and rankings on a Hacker News dataset, and ending with a summary of session‑level metrics (requests, tokens, cost, cache hits). I also linked the GitHub repo and a hands‑on guide for building a Jev‑style model locally.
These posts reinforce my earlier decision to embed Jev directly in the database rather than route rows through external application code, because it eliminates data movement, keeps the schema simple (no vector columns or embedding indexes), and lets me express rich semantic queries directly in SQL.
I demonstrated pg-jev, an open‑source PostgreSQL extension that exposes the Jev LLM through SQL functions, letting me run semantic filters, classifications, and rankings directly in SQL without any vector column or embedding index.
pg-jev is an open-source Postgres extension that exposes Jev through SQL functions.
I chose pg-jev over building a separate vector store because it eliminates the overhead of maintaining embeddings and an ANN index while still enabling semantic queries directly in SQL. I also batch rows before sending them to Jev to reduce API calls and latency.