Skip to content

feat(search): Cosmos full-text search as composable index mode #180

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
prsasattms merged 1 commit into from
Jun 4, 2026
Merged

feat(search): Cosmos full-text search as composable index mode #180

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
45 changes: 44 additions & 1 deletion search/schemas.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,6 +8,7 @@
"""
from __future__ import annotations

import re
from typing import Any, Dict, List, Literal, Optional, Union

from pydantic import BaseModel, Field, model_validator
Expand Down Expand Up @@ -129,17 +130,59 @@ class IndexFilter(BaseModel):
# -----------------------------------------------------------------------------


_FTS_FIELD_RE = re.compile(r"^[A-Za-z_][A-Za-z0-9_]*(\.[A-Za-z_][A-Za-z0-9_]*)*$")


class IndexSpec(BaseModel):
id: str = Field(..., description="Caller-supplied label; echoed in results")
store: StoreConfig = Field(..., discriminator="type")
mode: Literal["vector", "fts"] = Field(
default="vector",
description=(
"Search mode. 'vector' (default) performs ANN over a vector column. "
"'fts' performs Cosmos DB full-text search via FullTextScore and requires "
"the container to have a full-text indexing policy on `fts_field`."
),
)
vector: VectorConfig = Field(default_factory=VectorConfig)
embedding: EmbeddingPolicy = Field(..., discriminator="policy")
embedding: Optional[EmbeddingPolicy] = Field(
default=None,
discriminator="policy",
description="Required when mode='vector'; ignored when mode='fts'.",
)
fts_field: Optional[str] = Field(
default=None,
description=(
"Document path to full-text search (e.g. 'content' or 'body.text'). "
"When omitted, defaults to the first entry in `content_fields`. "
"Only used when mode='fts'."
),
)
content_fields: List[str] = Field(default_factory=lambda: ["content"])
return_fields: List[str] = Field(default_factory=list)
filter: Optional[IndexFilter] = None
top_k: Optional[int] = Field(default=None, ge=1, le=500)
pipeline_id: Optional[str] = Field(default=None, description="Source pipeline id; surfaced into result metadata for blob preview")

@model_validator(mode="after")
def _validate_mode(self):
if self.mode == "vector":
if self.embedding is None:
raise ValueError("embedding is required when mode='vector'")
else: # fts
if not isinstance(self.store, CosmosStore):
raise ValueError("mode='fts' is only supported for cosmosdb store in v1")
field = self.fts_field or (self.content_fields[0] if self.content_fields else None)
if not field:
raise ValueError(
"mode='fts' requires `fts_field` or at least one entry in `content_fields`"
)
if not _FTS_FIELD_RE.match(field):
raise ValueError(
f"invalid fts_field {field!r}: must match [A-Za-z_][A-Za-z0-9_]*(\\.[A-Za-z_][A-Za-z0-9_]*)*"
)
return self


# -----------------------------------------------------------------------------
# Merge + include
Expand Down
193 changes: 191 additions & 2 deletions search/searcher.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -278,6 +278,126 @@ def _run():
return await asyncio.to_thread(_run)


# =============================================================================
# Cosmos full-text search
# =============================================================================


_FTS_FIELD_RE = re.compile(r"^[A-Za-z_][A-Za-z0-9_]*(\.[A-Za-z_][A-Za-z0-9_]*)*$")
_FTS_TOKEN_RE = re.compile(r"[A-Za-z0-9_'\-]+")
_FTS_MAX_TERMS = int(os.getenv("SEARCH_FTS_MAX_TERMS", "16"))


def _tokenize_fts_query(query: str) -> List[str]:
"""Conservative tokenizer for Cosmos FullTextScore.

Splits on non-alphanumeric (keeping hyphens, apostrophes, underscores),
lowercases, drops empties + duplicates while preserving order, and caps
at SEARCH_FTS_MAX_TERMS to bound SQL size + RU cost.
"""
if not query:
return []
seen: Dict[str, None] = {}
for tok in _FTS_TOKEN_RE.findall(query.lower()):
if tok and tok not in seen:
seen[tok] = None
if len(seen) >= _FTS_MAX_TERMS:
break
return list(seen.keys())


async def search_cosmos_fts(
store: CosmosStore,
fts_field: str,
query_terms: List[str],
top_k: int,
content_fields: List[str],
return_fields: List[str],
index_filter: Optional[IndexFilter],
include_vector: bool,
) -> List[dict]:
"""Cosmos DB full-text search via `FullTextScore` (sync SDK wrapped).

Requires the target container to have a full-text indexing policy on
`fts_field`. FullTextScore can only appear in `ORDER BY RANK`, so the
native BM25 score is not returned — hits get `score=None` and merge via
rank-derived RRF.
"""
from azure.cosmos import CosmosClient
from azure.identity import DefaultAzureCredential

if not _FTS_FIELD_RE.match(fts_field):
raise ValueError(f"invalid fts_field {fts_field!r}")
if not query_terms:
return []

if getattr(store.auth, "mode", "managed_identity") == "key":
key = await resolve_secret_ref(store.auth.secret_ref) # type: ignore[attr-defined]
client_args = {"credential": key}
else:
client_args = {"credential": DefaultAzureCredential()}

def _run():
client = CosmosClient(store.endpoint, **client_args)
database = client.get_database_client(store.database)
container = database.get_container_client(store.container)

ff = fts_field.lstrip("/").replace("/", ".")

params: List[dict] = [{"name": "@top_k", "value": int(top_k)}]
term_placeholders: List[str] = []
for i, term in enumerate(query_terms):
ph = f"@term{i}"
term_placeholders.append(ph)
params.append({"name": ph, "value": term})

where = ""
if index_filter and index_filter.where:
where = f"WHERE {index_filter.where}"
if isinstance(index_filter.params, dict):
for k, v in index_filter.params.items():
name = k if k.startswith("@") else f"@{k}"
params.append({"name": name, "value": v})

term_args = ", ".join(term_placeholders)
query = (
f"SELECT TOP @top_k c FROM c {where} "
f"ORDER BY RANK FullTextScore(c.{ff}, {term_args})"
)

hits: List[dict] = []
for item in container.query_items(
query=query, parameters=params, enable_cross_partition_query=True
):
if isinstance(item, dict) and "c" in item:
doc = item.get("c") or {}
else:
# Cosmos may return the doc inlined when only one column is projected.
doc = item if isinstance(item, dict) else {}
text, text_parts = _extract_text(doc, content_fields)

metadata: Dict[str, Any] = dict(doc.get("metadata", {}) or {})
for f in return_fields:
if f in doc and f not in metadata:
metadata[f] = doc[f]

hit: Dict[str, Any] = {
"id": doc.get("id"),
"score": None, # FullTextScore not available in projection
"text": text,
"text_parts": text_parts,
"metadata": metadata,
"source": doc.get("source"),
"source_ref": doc.get("source_ref") or doc.get("title") or doc.get("url"),
}
if include_vector:
hit["vector"] = None
hits.append(hit)
return hits

return await asyncio.to_thread(_run)


async def search_pgvector(
store: PgVectorStore,
vec_field: str,
Expand Down Expand Up @@ -401,6 +521,50 @@ async def _search_one_index(
info = PerIndexInfo(index_id=idx.id)
per_top_k = idx.top_k or default_per_index_top_k

# -------------------------------------------------------------------------
# FTS branch — no embedding call; Cosmos-only in v1.
# -------------------------------------------------------------------------
if idx.mode == "fts":
info.embedding_model = "fts"
if not query:
info.error = "fts mode requires text query"
return info, []
if not isinstance(idx.store, CosmosStore):
info.error = "fts mode only supports cosmosdb store in v1"
return info, []
terms = _tokenize_fts_query(query)
if not terms:
info.error = "fts mode: query produced no searchable tokens"
return info, []
fts_field = idx.fts_field or (idx.content_fields[0] if idx.content_fields else "content")
t1 = time.time()
try:
hits = await asyncio.wait_for(
search_cosmos_fts(
idx.store, fts_field, terms, per_top_k,
idx.content_fields, idx.return_fields, idx.filter, include_vector,
),
timeout=PER_INDEX_TIMEOUT_S,
)
except asyncio.TimeoutError:
info.error = f"per-index timeout > {PER_INDEX_TIMEOUT_S}s"
info.search_ms = int((time.time() - t1) * 1000)
return info, []
except Exception as e:
info.error = f"search: {str(e)[:180]}"
info.search_ms = int((time.time() - t1) * 1000)
return info, []
info.search_ms = int((time.time() - t1) * 1000)
info.result_count = len(hits)
return info, hits

# -------------------------------------------------------------------------
# Vector branch — embed then search.
# -------------------------------------------------------------------------
if idx.embedding is None:
info.error = "embedding policy required for vector mode"
return info, []

# Embed
t0 = time.time()
try:
Expand Down Expand Up @@ -573,6 +737,10 @@ async def run_search(http: httpx.AsyncClient, req: SearchRequest) -> SearchRespo
warnings.append(
"mixed embedding models with merge.strategy=score; scores are not directly comparable — consider 'rrf'"
)
if req.merge.strategy == "score" and any(ix.mode == "fts" for ix in req.indexes):
warnings.append(
"merge.strategy=score with fts indexes will rank fts hits as None — use 'rrf' for hybrid"
)

if req.strict and errors:
raise RuntimeError(
Expand All @@ -587,7 +755,10 @@ async def run_search(http: httpx.AsyncClient, req: SearchRequest) -> SearchRespo
idx_modality: Dict[str, str] = {}
idx_pipeline: Dict[str, str] = {}
for ix in req.indexes:
mod = getattr(ix.embedding, "input_modality", "text")
if ix.mode == "fts":
mod = "text"
else:
mod = getattr(ix.embedding, "input_modality", "text") if ix.embedding else "text"
idx_modality[ix.id] = mod
pid = getattr(ix, "pipeline_id", None)
if pid:
Expand Down Expand Up @@ -646,15 +817,33 @@ async def explain_search(req: SearchRequest) -> Dict[str, Any]:
"""Dry-run: return the resolved plan without executing embedding/search."""
plan = []
for idx in req.indexes:
if idx.mode == "fts":
fts_field = idx.fts_field or (idx.content_fields[0] if idx.content_fields else None)
plan.append({
"index_id": idx.id,
"mode": "fts",
"store_type": idx.store.type,
"fts_field": fts_field,
"fts_terms": _tokenize_fts_query(req.query or ""),
"embedding": None,
"content_fields": idx.content_fields,
"return_fields": idx.return_fields,
"has_filter": idx.filter is not None,
"top_k": idx.top_k or req.merge.per_index_top_k,
})
continue
emb = idx.embedding
if isinstance(emb, ModelEmbedding):
emb_desc = {"policy": "model", "model_id": emb.model_id, "normalize": emb.normalize}
elif isinstance(emb, PipelineEmbedding):
emb_desc = {"policy": "pipeline", "pipeline": emb.pipeline, "normalize": emb.normalize}
else:
elif isinstance(emb, PrecomputedEmbedding):
emb_desc = {"policy": "precomputed", "dims": len(emb.vector)}
else:
emb_desc = None
plan.append({
"index_id": idx.id,
"mode": "vector",
"store_type": idx.store.type,
"vector_field": idx.vector.field,
"metric": idx.vector.metric,
Expand Down
59 changes: 44 additions & 15 deletions tests/unit/__snapshots__/test_openapi_snapshots.ambr
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9097,25 +9097,33 @@
'type': 'array',
}),
'embedding': dict({
'discriminator': dict({
'mapping': dict({
'model': '#/components/schemas/ModelEmbedding',
'pipeline': '#/components/schemas/PipelineEmbedding',
'precomputed': '#/components/schemas/PrecomputedEmbedding',
}),
'propertyName': 'policy',
}),
'oneOf': list([
dict({
'$ref': '#/components/schemas/ModelEmbedding',
}),
'anyOf': list([
dict({
'$ref': '#/components/schemas/PipelineEmbedding',
'discriminator': dict({
'mapping': dict({
'model': '#/components/schemas/ModelEmbedding',
'pipeline': '#/components/schemas/PipelineEmbedding',
'precomputed': '#/components/schemas/PrecomputedEmbedding',
}),
'propertyName': 'policy',
}),
'oneOf': list([
dict({
'$ref': '#/components/schemas/ModelEmbedding',
}),
dict({
'$ref': '#/components/schemas/PipelineEmbedding',
}),
dict({
'$ref': '#/components/schemas/PrecomputedEmbedding',
}),
]),
}),
dict({
'$ref': '#/components/schemas/PrecomputedEmbedding',
'type': 'null',
}),
]),
'description': "Required when mode='vector'; ignored when mode='fts'.",
'title': 'Embedding',
}),
'filter': dict({
Expand All @@ -9128,11 +9136,33 @@
}),
]),
}),
'fts_field': dict({
'anyOf': list([
dict({
'type': 'string',
}),
dict({
'type': 'null',
}),
]),
'description': "Document path to full-text search (e.g. 'content' or 'body.text'). When omitted, defaults to the first entry in `content_fields`. Only used when mode='fts'.",
'title': 'Fts Field',
}),
'id': dict({
'description': 'Caller-supplied label; echoed in results',
'title': 'Id',
'type': 'string',
}),
'mode': dict({
'default': 'vector',
'description': "Search mode. 'vector' (default) performs ANN over a vector column. 'fts' performs Cosmos DB full-text search via FullTextScore and requires the container to have a full-text indexing policy on `fts_field`.",
'enum': list([
'vector',
'fts',
]),
'title': 'Mode',
'type': 'string',
}),
'pipeline_id': dict({
'anyOf': list([
dict({
Expand Down Expand Up @@ -9190,7 +9220,6 @@
'required': list([
'id',
'store',
'embedding',
]),
'title': 'IndexSpec',
'type': 'object',
Expand Down
Loading
Loading