Company-brain graph outperforms retrieval on HERB held-out company questions

·5 views

A developer built inherit-brain for a Graph Hacks track on company knowledge, storing each product's records as a FalkorDB graph of people, documents, bugs, and pull requests. An agent answers questions by calling nine read-only tools that run fixed, parameterised Cypher queries rather than writing its own. On a spec question for one product, the graph tools returned all ten authors and reviewers, while retrieval with the same model returned only one. The project was tested on Salesforce's HERB benchmark, a simulated company, using 30 product graphs and 1,203 held-out questions. The author reports that a complex bug-reopening question was answered correctly through a multi-step join, where retrieval named the wrong engineers.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Anchor Coverage: A CI Gate That Checks Model-Drafted Docs Against Evidence

A DEV Community post describes a CI gate for documentation changes where model-drafted prose must be traceable to machine-readable sources. The author illustrates the problem with a hypothetical docs PR containing a renamed CLI flag that linting never caught. The proposed workflow splits each change into three lanes: script-extracted facts, model-drafted connective prose confined to delimited blocks, and human-owned commitments the model may not touch. A Python script, anchor_gate.py, enforces this by requiring every fact in a draft block to carry a claim anchor resolved against a claims file. The gate fails the build when the share of unanchored paragraphs exceeds a configured budget.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Launches Agent Colony, a Community Where Only AI Agents Can Post

A developer has built Agent Colony, a platform where AI agents alone can post, comment, claim tasks, and earn reputation, with humans limited to read-only access. The project aims to test whether agents can self-organize and collaborate without human moderation. Agents receive cryptographic Ed25519 identities and earn karma for completed tasks, which can be carried across platforms. The creator reports more than 100 registered agents, over 7,000 messages, and more than 10 open tasks with rewards. The developer says agents can self-organize to some degree but still need help getting started.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

ShilpiList Desk Uses Gemma to Generate Multilingual Listings for Odisha Craft Sellers

A developer has built ShilpiList Desk, an AI-assisted tool that helps small craft sellers in Odisha draft product listings in English, Hindi, and Odia. Sellers upload product photos and supply facts such as product name, craft technique, material, dimensions, maker, and GI status, and the app produces listing drafts. The tool is built on the open-weight Gemma 4 model, with FastAPI, SQLite, and built-in checks for malformed output and unsupported claims. In a small evaluation of six sample products, all generated valid listings, though Hindi and Odia glossary term hit rates were only 50%. The developer notes that Odia drafts require native-speaker review and that willingness to pay has not yet been validated.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

REA: Toolkit Open-Source untuk Membedah dan Menganalisis Aplikasi Modern

REA (Reverse Engineer Anything) is an open-source framework designed to help developers understand and extract the architecture of software applications without manually reading code line by line. It combines AST-based static analysis, API and network contract extraction, and protocol handshake reconstruction into a modular analysis pipeline. The tool can process targets such as bundled JavaScript files, Node.js applications, and Electron desktop apps, producing architectural reports. It addresses common challenges like incomplete documentation and obfuscated, minified bundles that make manual reverse engineering slow and tedious.

0 comments Read more at DEV Community