Knowledge / Featured system
FeedSilo
A durable research system that turns noisy public feeds into multilingual, multimodal, searchable knowledge.
01 / Context
Context
Research arrives as fragmented posts, threads, videos, feeds, pages, screenshots, and media. Saving a link is not enough when the underlying material can disappear, change, or require translation and multimodal interpretation.
02 / Problem
Problem
The system has to capture provenance-rich source material, preserve media, constrain model enrichment, protect technical tokens during translation, recover from queue failures, and still give a person control over what enters a durable knowledge base.
03 / System
System
FeedSilo ingests X, RSS, YouTube, web, manual, and Hermes-agent inputs; preserves raw context and media; runs structured translation and multimodal enrichment; embeds the result; and exposes hybrid retrieval with explicit review and routing states.
04 / Architecture
See the connective layer.
＋Architecture mode
System architecture / FeedSilo
Source-specific capture feeds a durable normalization and media queue before enrichment, embedding, indexing, human review, and retrieval.
- Public feed, page, video, or deliberate submission
- Forensic capture and raw hash
- Durable media queue
- Structured translation and multimodal enrichment
- Embedding and lexical indexes
- Human review and search
05 / Hard parts
Where the work actually was
- 01
Capturing network-level context and media without reducing a thread or article to a bare URL.
- 02
Preserving code, proper names, handles, URLs, and model identifiers during multilingual translation.
- 03
Recovering enrichment backlogs without duplicating work or silently losing failed items.
- 04
Combining semantic and lexical search while retaining visible human review and routing states.
06 / What changed or was learned
The field note
A durable knowledge product needs preservation and recovery before clever retrieval. Once raw source context, media, queue state, and review boundaries are dependable, model enrichment becomes useful rather than authoritative.
07 / Evidence
Proof
Safe public evidence, translated from tests and runtime validation rather than raw internal logs.
Browser-extension capture preserves network payloads, threads, articles, media, engagement context, and raw hashes for reviewable provenance.
Media preservation, translation, analysis, prompt extraction, and embeddings run as recoverable jobs with visible review state.
Codex app-server tasks use strict structured outputs and preserve code, names, URLs, handles, and model identifiers during translation.
Search combines 4,096-dimensional embeddings with lexical retrieval and human-defined research, creative, or mixed routing.
08 / Stack
Technology in service of the system
- Next.js
- React
- TypeScript
- Prisma
- PostgreSQL
- pgvector
- Docker
- Queue workers
- Codex app-server
- Vitest
- Playwright
10 / Status and boundaries