Qdrant: vector-store nIvbogh, qaStaHvIS reH Qap
Qu'

Qdrant: vector-store nIvbogh, qaStaHvIS reH Qap

Docker lo'laHghach Qdrant yIchu', De' (Dokumente) yInobtaH, qech ('Bedeutung') yISam — 'ej: nuq DaghajnIS embedding-qun, 'ej nuq DanobnIS.

mu'mey rammoH'a'?→ mu'tay'Daq yISam

latlh ghItlhDaq (vektorstore De') wIDellu': jan ('Werkzeug') 'oH, qech ('Bedeutung') Sammeh ngutlhmey ('Texte') SammoHbogh. QaQ theorie. DaHjaj wa' wIchu' — 'ej teH De'mey ('echte Dokumente') wInobtaH.

jan pong 'oH Qdrant'e' (jatlh „quadrant"). fonts-poSmoH ('Open Source') ghaH, machaSeylIj ('deine Maschine') Daq Qap, 'ej taghpu' QaQqu' — vIt jIjatlh.

nuq 'oH Qdrant'e' — 'ej nuq 'oHbe'

vector-store 'oH Qdrant'e': mIw ('Datenbank') 'oH, mI' vestlu' ('Embeddings') polbogh, 'ej nom rurbogh Sambogh. ngoDvam neH. cogh be', ja'chuq be', ngutlh ghItlh be'. Sam. QaQqu' 'ej nom.

ngoDvam Datlhojnis, 'ej puS janmey rur lurchugh. Datlhojbe'. OpenWebUI Hoch-De' polmeH jan ghaj. 'ach vector-store ghajbe' Flowise — 'ach wa' SeH-tlhegh ('Mausklick') lo'laHghach Qdrant rur jan luchel laH. vaj Qdrant 'oH stratum bIngDaq — wa' Daghajbogh, naDev Hoch DaSeHnIS.

chu'meH: wa' ra'wI'

Docker Qaprup, wa' tlhegh Qdrant Qap:

docker run -p 6333:6333 -p 6334:6334 \
  -v "$(pwd)/qdrant_storage:/qdrant/storage" \
  qdrant/qdrant

ngoDvam neH. -v pat De'lIj polmeH neH — chu'qa'DI', De' taHtaH (vumbe'chugh, vasDaq ratlh 'ej veb-vatlhvaD chIlbe').

navigatorDaq http://localhost:6333/dashboard Daq mach-mIllogh Daghaj, naDev collectionlIj Dalegh. ratlh be', noch be', ghItlh be'.

(qech mach: tessera Hutlh Qdrant taghDI'. machaSey-laghDaq experimentvaD QaQ — 'ach internet poSDaq Daqemchugh, security-paq yIlaD.)

De' yInob — exemplum

DaH 'ay' 'oH bIngDaq qaSbogh 'angbogh. ngutlh wIpe', Hoch frusta embedding-qunDaq wIngeH, Qdrant De' wIpol. Python Daq:

from qdrant_client import QdrantClient, models
from openai import OpenAI

qdrant = QdrantClient(url="http://localhost:6333")
ki = OpenAI()  # ambitusvo' clave laD

qdrant.recreate_collection(            # Qu' 1: collection chenmoH
    "ghItlhwIj",
    vectors_config=models.VectorParams(size=1536, distance=models.Distance.COSINE),
)  # 1536 = qun mI' tlhegh tIq

frusta = ["wa'DIch paragraph ...", "cha'DIch paragraph ...", "..."]
for i, ngutlh in enumerate(frusta):    # Qu' 2: yIngu' 'ej yIpol
    vector = ki.embeddings.create(
        model="text-embedding-3-small", input=ngutlh
    ).data[0].embedding
    qdrant.upsert("ghItlhwIj", [
        models.PointStruct(id=i, vector=vector, payload={"ngutlh": ngutlh})
    ])

yu' = ki.embeddings.create(            # Qu' 3: qech yISam
    model="text-embedding-3-small", input="nuq wIqelpu'?"
).data[0].embedding
SammeH = qdrant.query_points("ghItlhwIj", query=yu', limit=3)
for S in SammeH.points:
    print(S.payload["ngutlh"])

wej Qu': collection chenmoH, frusta ngu', Sam. vector-store paq mIw rur — 'ach DaH teH-mIw bIngDaq.

'ej DaH vIt: scriptvam DaghItlhnISbe' pIj. OpenWebUI insertion ghaj; FlowiseDaq Qdrant-bom ('Baustein') Dachel — 'ej cha' jan bIngDaq vum. exemplum 'oH bIngDaq qaSbogh DatlhojmeH neH. ghaytan Daghaj Hoch SeHwI'.

nuq DaghajnIS — KI 'ay'

embeddings pol Qdrant — 'ach embedding-qun ('Einbettungs-Modell') chenmoHnIS. KI Segh pagh 'oH: ngutlh mI'Daq mughmeH ngeb ('spezialisiert'), ja'chuqmeH be'. loS Hem:

  • OpenAI — text-embedding-3-small 'oH norm quietqu'. clave yISuq, Qap.
  • Jina AI — taHtaH, yIntaH 'ej tInchoHtaH. BerolinDaq ('Berlin') Daq, EuropaDaq De' vum, 'ej AVV (Auftragsverarbeitungsvertrag) chellaH — vaj DSGVO yu'vaD QaQ Europa jang. ('a DaH Elastic US-yejHaD Jina ghaj; 'ach EuropaDaq ratlh mIqta'mey, jatlh provisor.) taghmeH ratlhbe' portion tu'lu' — DaH wa' SaD SaD ('eine Million') token, mercatur-Hutlh.
  • OpenRouter — DaH embeddings je ghaj exemplar-boqHa'wI', wa' clave (OpenAI, Mistral, Qwen, Google). comparemeH QaQ.
  • Hoch local — Ollama lo'lu'chugh, machaSeyDaq embedding je Qap. yIqIm: ja'chuqwI'-qunlIj 'oHbe' — embedding Qu'vaD ngeb qun pIm 'oH ('eigenes Modell'). QaQ: nomic-embed-text (mach 'ej nom) ghap bge-m3 (HochHom Hol Sov, DoyIchlan HolvaD QaQ). clave be', noch be', De'lIj juH lImbe'. 'ach yIqaw: OpenAI rurbe' vector tIq (nomic-embed-text 768, 1536 be') — vaj codexDaq collection size yIchoH. tardqu' 'a ratlhbe' 'ej pegh.

Diary-vam tlhegh Doq: poSmoHbogh varietas tu'lu'chugh, yIlegh — De' pegh DI'.

nuq Daq embedding?

jang Dab: HochHom Daq mach. DaH text-embedding-3-small Daq cha' cent, wa' SaD SaD tokenvaD (token rur syllable).

qatlh? paq tIq SaDvatlh pagh inserevaD wa' cha' cent Daq pum. wa' yu' embedmeH Daq machqu' 'e' centDaq qonlaHbe'. 'ej local Ollama? pagh — machaSey 'ul neH.

intelligentia ration HochDaq embeddings Daq machqu'. naDev parc yImev.

cha' Hem lo'meH

Zoo Code — editorDaq codex-adiutor fonts-poSmoH, machaSey Qdrant qawHaq codexvaD lo'laH. Hoch yu'vaD codex-bas naQ ngeHbe', loQ Daq Sam neH. token pol 'ej jang chIm.

Flowise — naDev codex Hutlh Hoch KI-He chenmoH: De' yInob, qawHaq Qdrant, ja'chuqwI'-qun tlhop — 'ej ja'chuqwI' contentlIj Sovbogh tu'lu'. chovnatlh Daq paq pagh 'oH. veb 'oH.

DalughnIS'a'?

rut neH yu'chugh: ghobe', jan rIntaH Hoch-De' polmoHbogh yap. Qdrant DaSeHchugh DaH lugh, qech chu' Dachenmohrupchugh — ja'chuqwI', codex qaw, collection Sambogh, SoHvaD chenchoHbogh 'ej machaSey ratlhbogh.

Docker ra'wI' tagh QaQ. Qdrant yIchu', tabuh yIpoSmoH, yIbej. 'ej veb-vatlh Flowise wIlo' 'ej ja'chuqwI' wIchenmoH.