fleet-memory/hindsight-api/hindsight_api/engine/jina_mlx_reranker.py
Nicolò Boschi 1caf5ec9ee
feat: add jina-mlx reranker provider for Apple Silicon (#542)
* feat: add JinaMLXCrossEncoder for native Apple Silicon reranking

Adds a new `jina-mlx` reranker provider backed by jinaai/jina-reranker-v3-mlx,
a 0.6B multilingual listwise reranker running via the MLX framework on Apple Silicon.
The model is downloaded automatically from HuggingFace Hub on first use.

Benchmarked latencies (Apple Silicon): 1 doc→32ms, 5→45ms, 10→60ms, 20→94ms.
Sub-linear scaling because all docs are ranked in a single forward pass.

- Embeds the MLX reranker implementation (_MLXReranker / _MLPProjector) directly
  in cross_encoder.py with no transformers/PyTorch dependency
- Adds `mlx`, `mlx-lm`, `safetensors` to pyproject.toml optional deps (uv add)
- Updates configuration.md with provider docs and benchmark table

* refactor: import MLXReranker from repo rerank.py instead of duplicating code

Use importlib to load MLXReranker directly from the model repo's own rerank.py
(downloaded via snapshot_download). Also pin exact minimum versions for
mlx>=0.31.0, mlx-lm>=0.31.1, safetensors>=0.6.2 (verified against installed versions).

* refactor: move MLX reranker impl to dedicated jina_mlx_reranker.py

Replaces the importlib hack with a proper module. jina_mlx_reranker.py is
adapted from jinaai/jina-reranker-v3-mlx/rerank.py (CC BY-NC 4.0) with the
source clearly documented at the top of the file.

* docs: simplify jina-mlx reranker docs

* fix: disable GIN fastupdate on source_memory_ids index to prevent deadlocks

GIN fastupdate buffers inserts in a pending list and flushes it with
AccessExclusiveLock when full. Under concurrent test load (8 xdist workers
all running retain_async), two workers can trigger a flush simultaneously
and deadlock. Recreating the index with fastupdate=off eliminates the
flush/lock cycle at the cost of slightly slower individual inserts.

* fix: drop per-bank HNSW indexes after transaction to avoid AccessExclusiveLock deadlock

When deleting a bank, the previous code dropped HNSW indexes inside the
same transaction as the DELETE FROM memory_units. Since DROP INDEX needs
AccessExclusiveLock on the parent table and DELETE holds RowExclusiveLock,
two concurrent bank deletions deadlocked on the same table lock.

Fix: capture internal_id inside the transaction, commit, then drop the
indexes outside the transaction so no row-level locks are held.
2026-03-11 15:15:58 +01:00

144 lines
5.2 KiB
Python

"""
MLX implementation of jina-reranker-v3 for Apple Silicon.
This file is adapted from the official model repository:
https://huggingface.co/jinaai/jina-reranker-v3-mlx/blob/main/rerank.py
License: CC BY-NC 4.0 (contact Jina AI for commercial usage)
Changes from upstream:
- Removed the __main__ example block
- Type annotations added to public methods
- top_n parameter added to rerank() (upstream only exposed it implicitly)
"""
import numpy as np
class _MLPProjector:
def __init__(self):
import mlx.nn as nn
self.linear1 = nn.Linear(1024, 512, bias=False)
self.linear2 = nn.Linear(512, 512, bias=False)
def __call__(self, x):
import mlx.nn as nn
x = self.linear1(x)
x = nn.relu(x)
x = self.linear2(x)
return x
def _load_projector(projector_path: str) -> _MLPProjector:
import mlx.core as mx
from safetensors import safe_open
projector = _MLPProjector()
with safe_open(projector_path, framework="numpy") as f:
projector.linear1.weight = mx.array(f.get_tensor("linear1.weight"))
projector.linear2.weight = mx.array(f.get_tensor("linear2.weight"))
return projector
def _sanitize(text: str, special_tokens: dict[str, str]) -> str:
for token in special_tokens.values():
text = text.replace(token, "")
return text
def _format_prompt(query: str, docs: list[str], special_tokens: dict[str, str]) -> str:
query = _sanitize(query, special_tokens)
docs = [_sanitize(d, special_tokens) for d in docs]
doc_token = special_tokens["doc_embed_token"]
query_token = special_tokens["query_embed_token"]
prefix = (
"<|im_start|>system\n"
"You are a search relevance expert who can determine a ranking of the passages based on how relevant they are to the query. "
"If the query is a question, how relevant a passage is depends on how well it answers the question. "
"If not, try to analyze the intent of the query and assess how well each passage satisfies the intent. "
"If an instruction is provided, you should follow the instruction when determining the ranking."
"<|im_end|>\n<|im_start|>user\n"
)
suffix = "<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n"
body = (
f"I will provide you with {len(docs)} passages, each indicated by a numerical identifier. "
f"Rank the passages based on their relevance to query: {query}\n"
)
body += "\n".join(f'<passage id="{i}">\n{doc}{doc_token}\n</passage>' for i, doc in enumerate(docs))
body += f"\n<query>\n{query}{query_token}\n</query>"
return prefix + body + suffix
class MLXReranker:
"""
MLX-accelerated jina-reranker-v3 for Apple Silicon.
Loads the model from a local directory (use huggingface_hub.snapshot_download
to fetch jinaai/jina-reranker-v3-mlx if you don't have it already).
"""
_SPECIAL_TOKENS = {
"query_embed_token": "<|rerank_token|>",
"doc_embed_token": "<|embed_token|>",
}
_DOC_TOKEN_ID = 151670
_QUERY_TOKEN_ID = 151671
def __init__(self, model_path: str, projector_path: str):
from mlx_lm import load
self.model, self.tokenizer = load(model_path)
self.model.eval()
self.projector = _load_projector(projector_path)
def rerank(self, query: str, documents: list[str], top_n: int | None = None) -> list[dict]:
"""
Rank documents by relevance to a query.
Returns a list of dicts with keys: document, relevance_score, index.
Sorted by descending relevance_score.
"""
import mlx.core as mx
prompt = _format_prompt(query, documents, self._SPECIAL_TOKENS)
input_ids = self.tokenizer.encode(prompt)
hidden_states = self.model.model([input_ids])[0] # [seq_len, hidden_size]
input_ids_np = np.array(input_ids)
query_positions = np.where(input_ids_np == self._QUERY_TOKEN_ID)[0]
doc_positions = np.where(input_ids_np == self._DOC_TOKEN_ID)[0]
if len(query_positions) == 0:
raise ValueError("Query embed token not found in prompt")
if len(doc_positions) == 0:
raise ValueError("Document embed tokens not found in prompt")
query_hidden = mx.expand_dims(hidden_states[int(query_positions[0])], axis=0)
doc_hidden = mx.stack([hidden_states[int(p)] for p in doc_positions])
query_emb = self.projector(query_hidden) # [1, 512]
doc_emb = self.projector(doc_hidden) # [num_docs, 512]
query_exp = mx.broadcast_to(mx.expand_dims(query_emb, 0), (1, len(documents), 512))
doc_exp = mx.expand_dims(doc_emb, 0)
scores = mx.sum(doc_exp * query_exp, axis=-1) / (
mx.sqrt(mx.sum(doc_exp * doc_exp, axis=-1)) * mx.sqrt(mx.sum(query_exp * query_exp, axis=-1))
) # [1, num_docs]
scores_np = np.array(scores[0])
order = np.argsort(scores_np)[::-1]
n = min(top_n, len(documents)) if top_n is not None else len(documents)
return [
{
"document": documents[order[i]],
"relevance_score": float(scores_np[order[i]]),
"index": int(order[i]),
}
for i in range(n)
]