ingest_fragment

Benchmarked against: Anthropic — Memory tool MCP Tool: Available on both Cloud UB and Local UBI servers Cost: Free (internal operation)

ingest_fragment is the primary content ingestion tool — the entry point for all knowledge entering Universal Brain. It runs the full MTAAA pipeline to classify and store content.

---

Usage

# Ingest a file
ingest_fragment(path="/path/to/spec.md", input_type="file")

# Ingest text directly
ingest_fragment(path="LangGraph 1.0.9 released with new checkpoint API...", input_type="text")

# Ingest from URL
ingest_fragment(path="https://example.com/article", input_type="url")

Parameters

Parameter	Type	Default	Description
path	string	required	File path, text content, or URL
input_type	string	"file"	file / text / url / screenshot
source	string	"manual"	manual / api / nq_alpha / ss_vault / downloads

Response

{
  "entry_id": "ub-abc123def456",
  "category": "knowledge",
  "title": "LangGraph 1.0.9 Release Notes",
  "vectorized": true
}

The MTAAA pipeline

Every ingestion runs through the 5-node classification pipeline:

1. Caller → Agent calls ingest_fragment()
2. UBI Router → Detects content type, routes to handler
3. Handler (e.g., 文字鍋爺) → Classifies content
4. 3D Classification → Topic × Type × Lifecycle from Controlled Vocabulary
5. Result → Entry stored in D1 + vectorized for search

See File Ingestion (MTAAA) for full pipeline details.

Ingestion rules

From Company Constitution §1-§2 and UB Governance:

Rule	Detail
Language	All entries must be in English
Title	Descriptive, searchable
Content	Self-contained (reader needs no other context)
Tags	Lowercase, hyphenated, max 8 per entry
No duplicates	search_brain() first to check
Timestamp	Include Taipei time

Quality checklist

Before calling ingest_fragment():

1. Is the title descriptive and searchable?
2. Is the content self-contained?
3. Are tags following controlled vocabulary?
4. Did you search UB first to avoid duplicates?
5. Is the content in English?