endee 0.1.41b1

Endee is the Next-Generation Vector Database for Scalable, High-Performance AI

Endee - Python Client

Endee is a high-performance vector database with three retrieval modes you can mix per object and per query:

• Dense vector search - semantic ANN over embeddings (HNSW).
• Keyword search - sparse embeddings (BM25-style inverted index).
• Multi-vector search - many vectors per object, pooled for retrieval and reranked.

A single object can hold a dense field, a sparse field, and a multi-vector field at once, and a single query can hit any subset - returned per-field or fused with RRF.

Where it fits

• Edge / on-device - int16 / int8e precisions keep memory and footprint small and still provide 97% recall and 1500 QPS.
• E-commerce - semantic + keyword hybrid with metadata filters (price, category, …).
• High-performance agentic AI / RAG - fast multi-field retrieval and multi-vector reranking for low-latency tool/context lookup.

---

Concepts

Database                      ← top level; one database = one token, fully isolated
└── Collection                ← a named set of typed fields
    └── Object                ← one record
        ├── id                ← unique string key
        ├── meta              ← arbitrary JSON, returned with results
        ├── filter            ← tags used to filter searches
        └── fields            ← the vectors, by field name:
              embedding   (dense)         [0.1, 0.2, ...]
              keywords    (sparse)        {indices, values}
              multivec     (multi_vector)  [[...], [...]]

• Database - top-level namespace; created by the root token, addressed by a db token. Databases are isolated from each other.
• Collection - declares named, typed fields (vector / sparse / multi_vector) and holds objects.
• Object - one item in a collection:
  • id - unique string identifier (insert-or-replace key).
  • meta - arbitrary JSON payload, echoed back in search results.
  • filter - key/value tags queried with filter operators ($eq, $in, $range, …).
  • fields - the actual vectors, one entry per field you populate.

---

Install

pip install endee

Python ≥ 3.9. Depends on numpy, msgpack, orjson, requests, httpx.

---

Quick start

Authentication is always on - every request needs a token.

from endee import Endee

# 1) Admin: create a database with the ROOT token.
#    If the server was started WITHOUT NDD_ROOT_TOKEN, the root token is the
#    insecure dev default "unsafe_root_token".
admin = Endee(token="unsafe_root_token")
admin.set_base_url("http://localhost:8080/api/v2")
db_token = admin.create_database("alice")        # returns the db token, e.g. "alice:9f3..."

# 2) App: everything else uses the DB token.
client = Endee(token=db_token)
client.set_base_url("http://localhost:8080/api/v2")

# 3) Create a collection with a dense field.
client.create_collection(
    name="products",
    fields=[
        {"name": "embedding", "type": "vector",
         "params": {"dimension": 8, "space_type": "cosine", "precision": "int8"}},
    ],
)
collection = client.get_collection("products")

# 4) Insert objects (dense vector per object).
collection.upsert([
    {"id": "p1", "meta": {"name": "Wireless Headphones"}, "filter": {"category": "electronics"},
     "fields": {"embedding": [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8]}},
    {"id": "p2", "meta": {"name": "Running Shoes"}, "filter": {"category": "footwear"},
     "fields": {"embedding": [0.5, 0.4, 0.3, 0.2, 0.1, 0.0, 0.2, 0.3]}},
])

# 5) Search.
hits = collection.search(fields={"embedding": {"query": [0.2, 0.2, 0.3, 0.3, 0.4, 0.5, 0.6, 0.7], "limit": 3}})
for h in hits["results"]["embedding"]:
    print(h["id"], round(h["similarity"], 4), h["meta"])

Tokens: the root token (server's NDD_ROOT_TOKEN, or unsafe_root_token in dev) administers databases; a db token (db_name:secret, returned by create_database) does all data work. A missing/invalid token → 401; a read-only token on a write → 403.

---

Field types

Add fields at collection-creation time. Each is a plain dict:

Type	Keys	Value on upsert	query value
vector	params: dimension, space_type, precision	[0.1, 0.2, ...]	[0.2, ...]
sparse	sparse_model (top-level)	{"indices":[...], "values":[...]}	{"indices":[...], "values":[...]}
multi_vector	params: vector params + pooling ("mean"/"max")	[[...], [...]]	[[...], [...]]

On upsert, the value sits directly under the field name. On search, the query value is wrapped in a per-field config: {"query": <value>, "limit"?: ..., "ef_search"?: ...} (see Search).

space_type: cosine (default, normalized client-side), l2, ip. precision: float32, float16, int16, int8, int8e, binary.

Adding a sparse (keyword) field

client.create_collection("docs", fields=[
    {"name": "embedding", "type": "vector",
     "params": {"dimension": 8, "space_type": "cosine", "precision": "int8"}},
    {"name": "keywords", "type": "sparse", "sparse_model": "default"},
])

collection.upsert([{
    "id": "d1",
    "fields": {
        "embedding": [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8],
        "keywords":  {"indices": [3, 17, 42], "values": [0.9, 0.5, 0.2]},   # term ids + weights
    },
}])

Adding a multi-vector field

A multi-vector field stores several vectors per object, pooled into one vector for retrieval. Two pooling strategies are supported:

• mean (default) - elementwise average of the members.
• max - elementwise maximum of the members.

client.create_collection("passages", fields=[
    {"name": "embedding", "type": "vector",
     "params": {"dimension": 8, "space_type": "cosine", "precision": "int8"}},
    {"name": "multivec", "type": "multi_vector",
     "params": {"dimension": 8, "space_type": "cosine", "precision": "int8", "pooling": "mean"}},
])

collection.upsert([{
    "id": "x1",
    "fields": {
        "embedding": [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8],
        "multivec":   [[0.1]*8, [0.2]*8, [0.3]*8],   # one vector per token/passage
    },
}])

You can set any subset of fields on a given object; upsert is insert-or-replace by id.

---

Search

Every field is queried with the unified config {"query": ..., "limit"?: ..., "ef_search"?: ...} - query is required, and limit is the max hits to return for that field (default 10, max 4096). There's no overall result limit and no single-field special case: search always returns one ranked list per field, keyed by field name: {"results": {field: [hit, ...], ...}}. To search a single field, just pass that one field. A hit is {"id", "similarity", "meta", "filter"}; results carry meta/filter but not the stored vectors - use get_objects to fetch those.

Single field

res = collection.search(fields={"embedding": {"query": [0.2]*8, "limit": 5}})
res["results"]["embedding"]   # → [ {id, similarity, meta, filter}, ... ]

Multiple fields (per-field results)

Query several fields at once → each field's own ranked list (scores are per-field and not comparable across fields). Limits are per field:

res = collection.search(
    fields={
        "embedding": {"query": [0.2]*8, "limit": 50},
        "keywords":  {"query": {"indices": [3, 17], "values": [0.8, 0.4]}, "limit": 50},
    },
)
res["results"]["embedding"]   # → [ {id, similarity, ...}, ... ]
res["results"]["keywords"]    # → [ {id, similarity, ...}, ... ]

Fusing fields with RRF (rerank)

Reranking is a separate step: pass the search result into rerank to fuse the per-field lists into a single ranked list via Reciprocal Rank Fusion.

from endee import rerank

res = collection.search(fields={
    "embedding": {"query": [0.2]*8, "limit": 50},
    "keywords":  {"query": {"indices": [3, 17], "values": [0.8, 0.4]}, "limit": 50},
})
fused = rerank(
    res,
    limit=10,
    field_weights={"embedding": 0.6, "keywords": 0.4},   # optional, must sum to 1.0
    rrf_k=60,                                             # optional RRF smoothing constant k (default 60)
)
fused["results"]   # → [ {id, similarity (fused score), meta, filter}, ... ]

rrf_k is the RRF smoothing parameter k: a hit at rank r (1-based) in field f contributes field_weights[f] / (rrf_k + r) to its fused score. Larger k flattens the contribution gap between ranks (rank position matters less); smaller k sharpens it (top ranks dominate). Defaults to 60.

Multi-vector

collection.search(fields={"multivec": {"query": [[0.1]*8, [0.2]*8], "limit": 5}})
# multi-vector participates in multi-field search + rerank exactly like above.

Filters

filter is an array of conditions (AND-combined):

collection.search(
    fields={"embedding": {"query": [0.2]*8, "limit": 5}},
    filter=[{"category": {"$eq": "electronics"}}, {"price": {"$range": [10, 100]}}],
)

Operators: $eq, $in, $range, $gt, $gte, $lt, $lte.

Filter tuning (filtered search only) - two optional search params trade speed for recall when a filter is active:

collection.search(
    fields={"embedding": {"query": [0.2]*8, "limit": 5}},
    filter=[{"category": {"$eq": "electronics"}}],
    prefilter_cardinality_threshold=5000,   # ≤ this many matches → brute-force prefilter (1,000–1,000,000)
    filter_boost_percentage=25,             # expand HNSW candidate pool by 25% (0–100)
)

Return-shape summary

Step	results
search (any number of fields)	{field: [hit, ...], ...}
rerank(search_result, ...)	[hit, ...] (fused)

---

Object operations

collection.get_objects(["p1", "p2"])     # full objects: {id, meta, filter, vectors, sparses, multi_vectors}
collection.delete_object("p1")           # {"deleted": "p1"}
collection.delete_by_filter([{"category": {"$eq": "footwear"}}])   # {"deleted": <count>}
collection.update_filters([{"id": "p2", "filter": {"category": "sale"}}])   # {"updated": <count>}

Collection maintenance

collection.describe()                              # {name, fields, created_at, layout_version}

# Rebuild one or more dense fields' HNSW graphs in a single async call.
# Pass a list of field specs (server-native keys); only M/ef_con may change.
# Any sparse field in the list is skipped server-side.
collection.rebuild([
    {"field": "embedding", "M": 20, "ef_con": 200},
    {"field": "colbert",   "M": 12, "ef_con": 120},
])   # {"fields_total": 2, "total_objects": ..., "status": "in_progress"}

# Poll progress; reports the latest rebuilt field plus overall counts.
collection.rebuild_status()   # {field, fields_done, fields_total, percent_complete, status, ...}

collection.shrink()                                # reclaim disk
client.list_collections(); client.delete_collection("products")

Backups (per collection)

A backup captures one collection; restoring it creates a new collection. Backups are created on the collection; they're listed/restored/managed at the database level (the database owns its backups).

collection.create_backup("nightly")                    # back up THIS collection as a backup named "nightly" (async)
client.active_backup()                                  # status of the running backup; poll until it finishes
client.list_backups()                                   # list all backups in this database
client.backup_info("nightly")                           # metadata for the backup named "nightly"
client.restore_backup("nightly", "products_restored")   # restore backup "nightly" into a new collection "products_restored"
client.download_backup("nightly", "/tmp/nightly.tar")   # download backup "nightly" to the local file "/tmp/nightly.tar"
client.upload_backup("/tmp/nightly.tar")                # upload the local .tar file back into this database
client.delete_backup("nightly")                         # delete the backup named "nightly"

Token management

Admin (root token) - manage databases and their tokens:

admin.list_databases()                                       # list all databases
admin.get_database("alice")                                  # info for database "alice"
admin.deactivate_database("alice")                           # disable database "alice" (its tokens stop working; data kept)
admin.activate_database("alice")                             # re-enable database "alice"
admin.delete_database("alice")                               # delete database "alice" and ALL its data
admin.create_token("alice", name="ingest", token_type="rw")  # mint token "ingest" for db "alice"; returns the new db token (string)
admin.list_tokens("alice")                                   # list the token names/types of database "alice"
admin.delete_token("alice", "ingest")                        # delete the token named "ingest" from database "alice"
admin.list_db_collections("alice")                           # list collections in database "alice"
admin.list_all_collections()                                 # list collections across all databases

Self-service (db token) - a database manages its own tokens, no root needed:

client.create_my_token("worker", token_type="r")            # mint token "worker" for YOUR db; returns the new db token (string)
client.list_my_tokens()                                     # list your db's token names/types
client.delete_my_token("worker")                            # delete your db's token named "worker"

token_type: rw (read-write, default) or r (read-only - 403 on writes).

Server info

client.health()    # {"status": "ok", ...}
client.stats()     # {"version", "uptime", "total_requests"}

---

Reference

• Limits (client-side, fail fast): ≤ 10,000 objects per upsert; limit 1-4096; ef_search 1-1024; filter key ≤ 128 B, string value ≤ 1024 B; vector dim must match the field.
• Errors: non-2xx raise typed exceptions from endee - EndeeException (base), APIException, AuthenticationException (401), ForbiddenException (403), NotFoundException (404), ConflictException (409), ServerException (5xx). Bad inputs raise ValueError before any network call.

from endee import EndeeException, NotFoundException
try:
    collection.upsert(objects)
except NotFoundException as e:
    ...
except EndeeException as e:
    ...

---

Examples (scripts/)

• demo.py - full collection walkthrough. python3 scripts/demo.py --token alice:xxxx
• demo_admin.py - database administration. python3 scripts/demo_admin.py --root-token <root>
• test_filters.py - filter-operator checks. test_bulk_100k.py - bulk ingest + recall.