Five minutes to living context.

Install, create a database, add some vectors, link them into a graph, and query with context_chain. That's the whole loop.

The five-step loop

Create a database

Step 1 / 5

A Feather DB is just a file on disk. Open it (creating if missing) with a fixed dimension.

from feather_db import DB

# Create or open a 384-dimensional database
db = DB.open("embeddings.feather", dim=384)

Add vectors with metadata

Step 2 / 5

Every record carries an ID, a vector, and optional metadata — namespace, attributes, and importance.

import numpy as np
from feather_db import Metadata

for i in range(10):
    vec  = np.random.rand(384).astype(np.float32)
    meta = Metadata()
    meta.importance = 0.8
    meta.set_attribute("type", "brief")
    db.add(id=i, vec=vec, meta=meta)

print(f"Added {10} vectors")

Link records into a graph

Step 3 / 5

Typed edges turn your vectors into a traversable knowledge graph. Weighted, bidirectional.

# "informed_by" is a typed edge — you define what it means
db.link(from_id=0, to_id=1, rel_type="informed_by", weight=0.9)
db.link(from_id=0, to_id=2, rel_type="targets",     weight=0.7)

Query with context_chain

Step 4 / 5

Semantic search + BFS graph expansion in one call. Returns the full neighborhood, not just the nearest neighbors.

query = np.random.rand(384).astype(np.float32)

# k nearest, then 2 graph-hops of expansion
results = db.context_chain(query, k=5, hops=2)

for r in results:
    print(r.id, r.score, r.meta.get_attribute("type"))

Persist and reopen

Step 5 / 5

Feather writes a zero-copy binary file. Reopen it instantly from any process, any language binding.

# Flush state to disk
db.save()

# Reopen later (or from another process)
db = DB.open("embeddings.feather", dim=384)

Installation

API Reference