compair-core 0.3.12__tar.gz → 0.3.14__tar.gz

{compair_core-0.3.12 → compair_core-0.3.14}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compair-core
-Version: 0.3.12
+Version: 0.3.14
 Summary: Open-source foundation of the Compair collaboration platform.
 Author: RocketResearch, Inc.
 License: MIT
@@ -92,6 +92,8 @@ Key environment variables for the core edition:
 - `COMPAIR_REQUIRE_AUTHENTICATION` (`true`) – set to `false` to run the API in single-user mode without login or account management. When disabled, Compair auto-provisions a local user, group, and long-lived session token so you can upload documents immediately.
 - `COMPAIR_SINGLE_USER_USERNAME` / `COMPAIR_SINGLE_USER_NAME` – override the email-style username and display name that are used for the auto-provisioned local user in single-user mode.
 - `COMPAIR_INCLUDE_LEGACY_ROUTES` (`false`) – opt-in to the full legacy API surface (used by the hosted product) when running the core edition. Leave unset to expose only the streamlined single-user endpoints in Swagger.
+- `COMPAIR_EMBEDDING_DIM` – force the embedding vector size stored in the database (defaults to 384 for core, 1536 for cloud). Keep this in sync with whichever embedding model you configure.
+- `COMPAIR_VECTOR_BACKEND` (`auto`) – set to `pgvector` when running against PostgreSQL with the pgvector extension, or `json` to store embeddings as JSON (the default for SQLite deployments).
 
 See `compair_core/server/settings.py` for the full settings surface.
 

{compair_core-0.3.12 → compair_core-0.3.14}/README.md

@@ -57,6 +57,8 @@ Key environment variables for the core edition:
 - `COMPAIR_REQUIRE_AUTHENTICATION` (`true`) – set to `false` to run the API in single-user mode without login or account management. When disabled, Compair auto-provisions a local user, group, and long-lived session token so you can upload documents immediately.
 - `COMPAIR_SINGLE_USER_USERNAME` / `COMPAIR_SINGLE_USER_NAME` – override the email-style username and display name that are used for the auto-provisioned local user in single-user mode.
 - `COMPAIR_INCLUDE_LEGACY_ROUTES` (`false`) – opt-in to the full legacy API surface (used by the hosted product) when running the core edition. Leave unset to expose only the streamlined single-user endpoints in Swagger.
+- `COMPAIR_EMBEDDING_DIM` – force the embedding vector size stored in the database (defaults to 384 for core, 1536 for cloud). Keep this in sync with whichever embedding model you configure.
+- `COMPAIR_VECTOR_BACKEND` (`auto`) – set to `pgvector` when running against PostgreSQL with the pgvector extension, or `json` to store embeddings as JSON (the default for SQLite deployments).
 
 See `compair_core/server/settings.py` for the full settings surface.
 

{compair_core-0.3.12 → compair_core-0.3.14}/compair_core/api.py

@@ -1403,21 +1403,24 @@ def create_doc(
             datetime_modified=datetime.now(timezone.utc)
         )
         print('About to assign groups!')
-        print(groups.split(','))
-        if groups is not None:
-            group_ids = groups.split(',')
-            q = select(models.Group).filter(
-                models.Group.group_id.in_(group_ids)
-            )
-            groups = session.execute(q).fetchall()[0]
-            for group in groups:
-                document.groups.append(group)
+        target_group_ids = []
+        if groups:
+            target_group_ids = [gid.strip() for gid in groups.split(',') if gid.strip()]
+
+        if target_group_ids:
+            q = select(models.Group).filter(models.Group.group_id.in_(target_group_ids))
+            resolved_groups = session.execute(q).scalars().all()
+            if not resolved_groups:
+                raise HTTPException(status_code=404, detail="No matching groups found for provided IDs.")
+            document.groups = resolved_groups
         else:
-            q = select(models.Group).filter(
-                models.Group.name == current_user.username
-            )
-            group = session.execute(q).fetchone()[0]
-            document.groups = [group]
+            q = select(models.Group).filter(models.Group.name == current_user.username)
+            default_group = session.execute(q).scalars().first()
+            if default_group is None:
+                raise HTTPException(status_code=404, detail="Default group not found for user.")
+            document.groups = [default_group]
+
+        primary_group = document.groups[0]
 
         print(f'doc check!!! {document.content}')
         session.add(document)
@@ -1435,7 +1438,7 @@ def create_doc(
         log_activity(
             session=session, 
             user_id=document.author_id, 
-            group_id=group.group_id,
+            group_id=primary_group.group_id,
             action="create", 
             object_id=document.document_id, 
             object_name=document.title,

{compair_core-0.3.12 → compair_core-0.3.14}/compair_core/compair/embeddings.py

@@ -23,7 +23,17 @@ class Embedder:
 
         if self._cloud_impl is None:
             self.model = os.getenv("COMPAIR_LOCAL_EMBED_MODEL", "hash-embedding")
-            self.dimension = int(os.getenv("COMPAIR_LOCAL_EMBED_DIM", "384"))
+            default_dim = 1536 if self.edition == "cloud" else 384
+            dim_env = (
+                os.getenv("COMPAIR_EMBEDDING_DIM")
+                or os.getenv("COMPAIR_EMBEDDING_DIMENSION")
+                or os.getenv("COMPAIR_LOCAL_EMBED_DIM")
+                or str(default_dim)
+            )
+            try:
+                self.dimension = int(dim_env)
+            except ValueError:  # pragma: no cover - invalid configuration
+                self.dimension = default_dim
             base_url = os.getenv("COMPAIR_LOCAL_MODEL_URL", "http://local-model:9000")
             route = os.getenv("COMPAIR_LOCAL_EMBED_ROUTE", "/embed")
             self.endpoint = f"{base_url.rstrip('/')}{route}"

{compair_core-0.3.12 → compair_core-0.3.14}/compair_core/compair/main.py

@@ -12,7 +12,17 @@ from sqlalchemy.orm import Session as SASession
 
 from .embeddings import create_embedding, Embedder
 from .feedback import get_feedback, Reviewer
-from .models import Chunk, Document, Feedback, Group, Note, Reference, User
+from .models import (
+    Chunk,
+    Document,
+    Feedback,
+    Group,
+    Note,
+    Reference,
+    User,
+    VECTOR_BACKEND,
+    cosine_similarity,
+)
 from .utils import chunk_text, log_activity
 
 
@@ -159,22 +169,41 @@ def process_text(
         Chunk.note_id == note_id,
     ).first()
 
+    references: list[Chunk] = []
     if generate_feedback and existing_chunk:
         doc_group_ids = [g.group_id for g in doc.groups]
-        references = (
-            session.query(Chunk)
-            .join(Chunk.document)
-            .join(Document.groups)
-            .filter(
-                Document.is_published.is_(True),
-                Document.document_id != doc.document_id,
-                Chunk.chunk_type == "document",
-                Group.group_id.in_(doc_group_ids),
+        target_embedding = existing_chunk.embedding
+
+        if target_embedding is not None:
+            base_query = (
+                session.query(Chunk)
+                .join(Chunk.document)
+                .join(Document.groups)
+                .filter(
+                    Document.is_published.is_(True),
+                    Document.document_id != doc.document_id,
+                    Chunk.chunk_type == "document",
+                    Group.group_id.in_(doc_group_ids),
+                )
             )
-            .order_by(Chunk.embedding.cosine_distance(existing_chunk.embedding))
-            .limit(3)
-            .all()
-        )
+
+            if VECTOR_BACKEND == "pgvector":
+                references = (
+                    base_query.order_by(
+                        Chunk.embedding.cosine_distance(existing_chunk.embedding)
+                    )
+                    .limit(3)
+                    .all()
+                )
+            else:
+                candidates = base_query.all()
+                scored: list[tuple[float, Chunk]] = []
+                for candidate in candidates:
+                    score = cosine_similarity(candidate.embedding, target_embedding)
+                    if score is not None:
+                        scored.append((score, candidate))
+                scored.sort(key=lambda item: item[0], reverse=True)
+                references = [chunk for _, chunk in scored[:3]]
 
         sql_references: list[Reference] = []
         for ref_chunk in references:

{compair_core-0.3.12 → compair_core-0.3.14}/compair_core/compair/models.py

@@ -5,9 +5,15 @@ import hashlib
 import os
 import secrets
 from datetime import datetime, timezone
+from math import sqrt
+from typing import Sequence
 from uuid import uuid4
 
-from pgvector.sqlalchemy import Vector
+try:  # Optional: only required when using pgvector backend
+    from pgvector.sqlalchemy import Vector
+except ImportError:  # pragma: no cover - optional dependency in core
+    Vector = None  # type: ignore[assignment]
+
 from sqlalchemy import (
     Boolean,
     Column,
@@ -15,6 +21,7 @@ from sqlalchemy import (
     ForeignKey,
     Identity,
     Integer,
+    JSON,
     String,
     Table,
     Text,
@@ -27,6 +34,65 @@ from sqlalchemy.orm import (
     relationship,
 )
 
+_EDITION = os.getenv("COMPAIR_EDITION", "core").lower()
+_DEFAULT_DIM = 1536 if _EDITION == "cloud" else 384
+_DIM_ENV = (
+    os.getenv("COMPAIR_EMBEDDING_DIM")
+    or os.getenv("COMPAIR_EMBEDDING_DIMENSION")
+    or os.getenv("COMPAIR_LOCAL_EMBED_DIM")
+    or str(_DEFAULT_DIM)
+)
+
+try:
+    EMBEDDING_DIMENSION = int(_DIM_ENV)
+except ValueError:  # pragma: no cover - invalid configuration
+    EMBEDDING_DIMENSION = _DEFAULT_DIM
+
+
+def _detect_vector_backend() -> str:
+    explicit = os.getenv("COMPAIR_VECTOR_BACKEND")
+    if explicit:
+        return explicit.lower()
+
+    db = os.getenv("DB")
+    db_user = os.getenv("DB_USER")
+    db_passw = os.getenv("DB_PASSW")
+    db_url = os.getenv("DB_URL")
+    database_url = os.getenv("DATABASE_URL", "")
+
+    if all([db, db_user, db_passw, db_url]):
+        return "pgvector"
+    if database_url.lower().startswith(("postgres://", "postgresql://")):
+        return "pgvector"
+    return "json"
+
+
+VECTOR_BACKEND = _detect_vector_backend()
+
+
+def _embedding_column():
+    if VECTOR_BACKEND == "pgvector":
+        if Vector is None:
+            raise RuntimeError(
+                "pgvector is required when COMPAIR_VECTOR_BACKEND is set to 'pgvector'."
+            )
+        return mapped_column(Vector(EMBEDDING_DIMENSION), nullable=True)
+    # Store embeddings as JSON arrays (works across SQLite/Postgres without pgvector)
+    return mapped_column(JSON, nullable=True)
+
+
+def cosine_similarity(vec1: Sequence[float] | None, vec2: Sequence[float] | None) -> float | None:
+    if not vec1 or not vec2:
+        return None
+    if len(vec1) != len(vec2):
+        return None
+    dot = sum(a * b for a, b in zip(vec1, vec2))
+    norm1 = sqrt(sum(a * a for a in vec1))
+    norm2 = sqrt(sum(b * b for b in vec2))
+    if norm1 == 0 or norm2 == 0:
+        return None
+    return dot / (norm1 * norm2)
+
 
 class Base(DeclarativeBase, MappedAsDataclass):
     pass
@@ -216,7 +282,7 @@ class Document(BaseObject):
     file_key: Mapped[str | None] = mapped_column(String, nullable=True, default=None)
     image_key: Mapped[str | None] = mapped_column(String, nullable=True, default=None)
     is_published: Mapped[bool] = mapped_column(Boolean, default=False)
-    embedding = mapped_column(Vector(1536))
+    embedding: Mapped[list[float] | None] = _embedding_column()
 
     user = relationship("User", back_populates="documents")
     groups = relationship("Group", secondary="document_to_group", back_populates="documents")
@@ -250,7 +316,7 @@ class Note(Base):
     group_id: Mapped[str | None] = mapped_column(ForeignKey("group.group_id", ondelete="CASCADE"), index=True, nullable=True)
     content: Mapped[str] = mapped_column(Text)
     datetime_created: Mapped[datetime] = mapped_column(default=datetime.now(timezone.utc))
-    embedding = mapped_column(Vector(1536))
+    embedding: Mapped[list[float] | None] = _embedding_column()
 
     document = relationship("Document", back_populates="notes")
     author = relationship("User", back_populates="notes")
@@ -279,7 +345,7 @@ class Chunk(Base):
     document_id: Mapped[str | None] = mapped_column(ForeignKey("document.document_id", ondelete="CASCADE"), index=True, nullable=True)
     note_id: Mapped[str | None] = mapped_column(ForeignKey("note.note_id", ondelete="CASCADE"), index=True, nullable=True)
     chunk_type: Mapped[str] = mapped_column(String(16), default="document")
-    embedding = mapped_column(Vector(1536))
+    embedding: Mapped[list[float] | None] = _embedding_column()
 
     document = relationship("Document", back_populates="chunks")
     note = relationship("Note", back_populates="chunks")

{compair_core-0.3.12 → compair_core-0.3.14}/compair_core/server/local_model/app.py

@@ -2,6 +2,7 @@
 from __future__ import annotations
 
 import hashlib
+import os
 from typing import List
 
 from fastapi import FastAPI
@@ -9,7 +10,17 @@ from pydantic import BaseModel
 
 app = FastAPI(title="Compair Local Model", version="0.1.0")
 
-EMBED_DIMENSION = 384
+_DEFAULT_DIM = 384
+_DIM_ENV = (
+    os.getenv("COMPAIR_EMBEDDING_DIM")
+    or os.getenv("COMPAIR_EMBEDDING_DIMENSION")
+    or os.getenv("COMPAIR_LOCAL_EMBED_DIM")
+    or str(_DEFAULT_DIM)
+)
+try:
+    EMBED_DIMENSION = int(_DIM_ENV)
+except ValueError:  # pragma: no cover - invalid configuration
+    EMBED_DIMENSION = _DEFAULT_DIM
 
 
 def _hash_embedding(text: str, dimension: int = EMBED_DIMENSION) -> List[float]:

{compair_core-0.3.12 → compair_core-0.3.14}/compair_core.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compair-core
-Version: 0.3.12
+Version: 0.3.14
 Summary: Open-source foundation of the Compair collaboration platform.
 Author: RocketResearch, Inc.
 License: MIT
@@ -92,6 +92,8 @@ Key environment variables for the core edition:
 - `COMPAIR_REQUIRE_AUTHENTICATION` (`true`) – set to `false` to run the API in single-user mode without login or account management. When disabled, Compair auto-provisions a local user, group, and long-lived session token so you can upload documents immediately.
 - `COMPAIR_SINGLE_USER_USERNAME` / `COMPAIR_SINGLE_USER_NAME` – override the email-style username and display name that are used for the auto-provisioned local user in single-user mode.
 - `COMPAIR_INCLUDE_LEGACY_ROUTES` (`false`) – opt-in to the full legacy API surface (used by the hosted product) when running the core edition. Leave unset to expose only the streamlined single-user endpoints in Swagger.
+- `COMPAIR_EMBEDDING_DIM` – force the embedding vector size stored in the database (defaults to 384 for core, 1536 for cloud). Keep this in sync with whichever embedding model you configure.
+- `COMPAIR_VECTOR_BACKEND` (`auto`) – set to `pgvector` when running against PostgreSQL with the pgvector extension, or `json` to store embeddings as JSON (the default for SQLite deployments).
 
 See `compair_core/server/settings.py` for the full settings surface.
 

{compair_core-0.3.12 → compair_core-0.3.14}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "compair-core"
-version = "0.3.12"
+version = "0.3.14"
 description = "Open-source foundation of the Compair collaboration platform."
 readme = "README.md"
 license = { text = "MIT" }