Skip to content

This content is still in French

The page itself is translated, but its detailed content is not yet. Hardware sheets, guides and catalogues are still shown in French: we would rather give you the original than a rough machine translation. View the French site

Interroger ses propres documents hors ligne Étape 4 sur 6

Indexer

python
# indexer.py
import sqlite3, sqlite_vec, requests, struct

OLLAMA = "http://localhost:11434"
MODELE = "zylonai/multilingual-e5-small"
DIMENSION = 384

def vecteur(texte):
    r = requests.post(f"{OLLAMA}/api/embeddings",
                      json={"model": MODELE, "prompt": texte}, timeout=120)
    return r.json()["embedding"]

def encoder(v):
    return struct.pack(f"{len(v)}f", *v)

base = sqlite3.connect("corpus.db")
base.enable_load_extension(True)
sqlite_vec.load(base)

base.execute("CREATE TABLE IF NOT EXISTS passages(id INTEGER PRIMARY KEY, source TEXT, texte TEXT)")
base.execute(f"CREATE VIRTUAL TABLE IF NOT EXISTS vecteurs USING vec0(id INTEGER PRIMARY KEY, v FLOAT[{DIMENSION}])")

for doc in docs:                      # docs vient de decouper.py
    cur = base.execute("INSERT INTO passages(source, texte) VALUES (?, ?)",
                       (doc["source"], doc["texte"]))
    base.execute("INSERT INTO vecteurs(id, v) VALUES (?, ?)",
                 (cur.lastrowid, encoder(vecteur(doc["texte"]))))

base.commit()
print("index terminé")

Le résultat tient dans un seul fichier corpus.db. Il se copie sur une clé USB, se sauvegarde, se duplique sur toutes les bornes du parc.