Skip to content

refactor: separate concerns in RerankService with characterization tests #126

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
chottokun merged 2 commits into from
Oct 18, 2026
Merged

refactor: separate concerns in RerankService with characterization tests #126

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
06235c2
refactor: separate concerns in RerankService with characterization tests
chottokun Oct 15, 2026
6d4336d
test(rerank): add adversarial index bounds validation and edge case t…
chottokun Oct 18, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
186 changes: 102 additions & 84 deletions src/app/services/rerank.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,6 +17,7 @@


def _calculate_rerank_tokens(model: Any, query: str, documents: List[str]) -> Usage:
"""Calculates token usage for reranking query and documents batch."""
with model.tokenizer_lock:
tokenizer = model.tokenizer
q_tokens = len(tokenizer.encode(query, add_special_tokens=False))
Expand Down Expand Up @@ -48,6 +49,7 @@ def _calculate_rerank_tokens(model: Any, query: str, documents: List[str]) -> Us
def _sort_and_format_rerank_results(
results: List[dict], top_n: Optional[int]
) -> List[RerankData]:
"""Sorts rerank results by score descending and returns RerankData list."""
if top_n is not None:
sorted_results = heapq.nlargest(
top_n, results, key=lambda x: (x["score"], -x["document"])
Expand All @@ -59,19 +61,18 @@ def _sort_and_format_rerank_results(


class RerankService(BaseRerankService):
"""
Default production implementation of BaseRerankService.
"""
"""Default production implementation of BaseRerankService."""

def __init__(
self,
proxy_to_tei_func: Optional[Any] = None,
model_loader: Optional[Any] = None,
):
) -> None:
self.proxy_to_tei_func = proxy_to_tei_func
self.model_loader = model_loader

async def create_rerank(self, request: RerankRequest) -> RerankResponse:
"""Processes a rerank request by delegating to TEI proxy, Logit Gate, or standard cross-encoder."""
import app.main as main_mod

if (
Expand All @@ -87,92 +88,109 @@ async def create_rerank(self, request: RerankRequest) -> RerankResponse:
proxy_func = getattr(main_mod, "_proxy_to_tei", self.proxy_to_tei_func)

if tei_url and proxy_func:
tei_results = await proxy_func(
tei_url,
"/rerank",
{"query": request.query, "texts": request.documents},
)

results = []
for item in tei_results:
idx = item["index"]
score = item["score"]
result_item = {"document": idx, "score": float(score)}
if request.return_documents:
result_item["text"] = request.documents[idx]
results.append(result_item)

response_data = _sort_and_format_rerank_results(results, request.top_n)
usage = Usage(prompt_tokens=0, total_tokens=0)
return RerankResponse(
query=request.query,
data=response_data,
model=request.model,
usage=usage,
)
return await self._process_tei_proxy(request, tei_url, proxy_func)

if request.model in LOGIT_GATE_MODELS:
model = get_validated_model(
request.model,
LOGIT_GATE_MODELS,
"logit gate",
loader=self.model_loader,
)
return self._process_logit_gate(request)

return self._process_standard_rerank(request)

async def _process_tei_proxy(
self, request: RerankRequest, tei_url: str, proxy_func: Any
) -> RerankResponse:
"""Proxies rerank evaluation to Text Embeddings Inference (TEI) server."""
tei_results = await proxy_func(
tei_url,
"/rerank",
{"query": request.query, "texts": request.documents},
)

gate_service = LogitGateService(model)
matcher = AsciiMatcher(min_token_len=LOGIT_GATE_ASCII_MIN_TOKEN_LEN)
gate_results = gate_service.predict_margins(
request.query, request.documents
)
results = []
num_docs = len(request.documents)
for item in tei_results:
idx = item["index"]
if (
isinstance(idx, bool)
or not isinstance(idx, int)
or idx < 0
or idx >= num_docs
):
raise HTTPException(
status_code=500,
detail=f"TEI Proxy returned invalid document index: {idx}",
)
score = item["score"]
result_item = {"document": idx, "score": float(score)}
if request.return_documents:
result_item["text"] = request.documents[idx]
results.append(result_item)

beta = (
0.0
if request.use_ascii_boost is False
else LOGIT_GATE_ASCII_BOOST_WEIGHT
)
containment_scores = matcher.score_documents(
request.query, request.documents
)
response_data = _sort_and_format_rerank_results(results, request.top_n)
usage = Usage(prompt_tokens=0, total_tokens=0)
return RerankResponse(
query=request.query,
data=response_data,
model=request.model,
usage=usage,
)

th = (
request.threshold
if request.threshold is not None
else LOGIT_GATE_THRESHOLD
)
results = []
for i in range(len(request.documents)):
delta_z_base = gate_results[i]["logit_margin"]
containment = containment_scores[i]
delta_z_final = delta_z_base + (beta * containment)
score = _sigmoid(delta_z_final)
passed = score >= th

if request.drop_failed and not passed:
continue

entropy = _binary_entropy(score)

result_item = {
"document": i,
"score": float(score),
"passed": bool(passed),
"logit_margin": float(delta_z_final),
"containment_score": float(containment),
"entropy": float(entropy),
}
if request.return_documents:
result_item["text"] = request.documents[i]
results.append(result_item)

response_data = _sort_and_format_rerank_results(results, request.top_n)
usage = _calculate_rerank_tokens(model, request.query, request.documents)
return RerankResponse(
query=request.query,
data=response_data,
model=request.model,
usage=usage,
)
def _process_logit_gate(self, request: RerankRequest) -> RerankResponse:
"""Processes rerank request using hybrid Logit Gate model and ASCII boost."""
model = get_validated_model(
request.model,
LOGIT_GATE_MODELS,
"logit gate",
loader=self.model_loader,
)

gate_service = LogitGateService(model)
matcher = AsciiMatcher(min_token_len=LOGIT_GATE_ASCII_MIN_TOKEN_LEN)
gate_results = gate_service.predict_margins(request.query, request.documents)

beta = (
0.0 if request.use_ascii_boost is False else LOGIT_GATE_ASCII_BOOST_WEIGHT
)
containment_scores = matcher.score_documents(request.query, request.documents)

th = (
request.threshold if request.threshold is not None else LOGIT_GATE_THRESHOLD
)
results = []
for i in range(len(request.documents)):
delta_z_base = gate_results[i]["logit_margin"]
containment = containment_scores[i]
delta_z_final = delta_z_base + (beta * containment)
score = _sigmoid(delta_z_final)
passed = score >= th

if request.drop_failed and not passed:
continue

entropy = _binary_entropy(score)

result_item = {
"document": i,
"score": float(score),
"passed": bool(passed),
"logit_margin": float(delta_z_final),
"containment_score": float(containment),
"entropy": float(entropy),
}
if request.return_documents:
result_item["text"] = request.documents[i]
results.append(result_item)

response_data = _sort_and_format_rerank_results(results, request.top_n)
usage = _calculate_rerank_tokens(model, request.query, request.documents)
return RerankResponse(
query=request.query,
data=response_data,
model=request.model,
usage=usage,
)

def _process_standard_rerank(self, request: RerankRequest) -> RerankResponse:
"""Processes rerank request using standard cross-encoder model."""
model = get_validated_model(
request.model,
RERANK_MODELS,
Expand Down
Loading
Loading