How Kioku turns raw data into searchable knowledge.

Embedding Model

Kioku uses nomic-embed-text-v2-moe via Ollama for all embeddings.

Metric	Value
MTEB score	63.9
Dimensions	256–768 (configurable)
Latency (GPU)	5–20ms
Latency (CPU)	50–200ms
Cost	Free (compute only)
Privacy	Data stays on your server

This model matches OpenAI's text-embedding-3-small (62.3 MTEB) on benchmarks while running entirely on your hardware.

Pipeline

Documents (PDF)

Upload PDF → pdf-extract (text) → chunk → Ollama embed → Qdrant store

1. PDF uploaded via POST /knowledge/documents
2. Text extracted using pdf-extract crate
3. Text split into chunks
4. Each chunk embedded via Ollama HTTP API
5. Embeddings + metadata stored in Qdrant

Meetings (Transcripts)

Meeting transcript → per-segment embed → Qdrant store

1. Transcript ingested via POST /meetings
2. Each transcript segment (speaker + text + timestamps) embedded
3. Embeddings + meeting metadata stored in Qdrant
4. Searchable via POST /knowledge/search

Search

Vector similarity search across all knowledge (documents + meetings):

curl -X POST http://localhost:9100/knowledge/search \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"query":"deployment strategy","limit":5}'

Results are ranked by semantic similarity score (0–1).