compair-core 0.3.14__py3-none-any.whl → 0.3.15__py3-none-any.whl

compair_core/compair/feedback.py

@@ -1,12 +1,18 @@
 from __future__ import annotations
 
 import os
+from typing import Any, Iterable, List
+
 import requests
-from typing import Any
 
 from .logger import log_event
 from .models import Document, User
 
+try:
+    import openai  # type: ignore
+except ImportError:  # pragma: no cover - optional dependency
+    openai = None  # type: ignore
+
 try:
     from compair_cloud.feedback import Reviewer as CloudReviewer  # type: ignore
     from compair_cloud.feedback import get_feedback as cloud_get_feedback  # type: ignore
@@ -16,64 +22,183 @@ except (ImportError, ModuleNotFoundError):
 
 
 class Reviewer:
-    """Edition-aware wrapper that falls back to the local feedback endpoint."""
+    """Edition-aware wrapper that selects a feedback provider based on configuration."""
 
     def __init__(self) -> None:
         self.edition = os.getenv("COMPAIR_EDITION", "core").lower()
+        self.provider = os.getenv("COMPAIR_GENERATION_PROVIDER", "local").lower()
+        self.length_map = {
+            "Brief": "1–2 short sentences",
+            "Detailed": "A couple short paragraphs",
+            "Verbose": "As thorough as reasonably possible without repeating information",
+        }
+
         self._cloud_impl = None
+        self._openai_client = None
+        self.openai_model = os.getenv("COMPAIR_OPENAI_MODEL", "gpt-4o-mini")
+
         if self.edition == "cloud" and CloudReviewer is not None:
             self._cloud_impl = CloudReviewer()
+            self.provider = "cloud"
         else:
-            self.client = None
-            self.model = os.getenv("COMPAIR_LOCAL_GENERATION_MODEL", "local-feedback")
-            base_url = os.getenv("COMPAIR_LOCAL_MODEL_URL", "http://local-model:9000")
-            route = os.getenv("COMPAIR_LOCAL_GENERATION_ROUTE", "/generate")
-            self.endpoint = f"{base_url.rstrip('/')}{route}"
+            if self.provider == "openai":
+                api_key = os.getenv("COMPAIR_OPENAI_API_KEY")
+                if api_key and openai is not None:
+                    # Support both legacy (ChatCompletion) and new SDKs
+                    if hasattr(openai, "api_key"):
+                        openai.api_key = api_key  # type: ignore[assignment]
+                    if hasattr(openai, "OpenAI"):
+                        try:  # pragma: no cover - optional runtime dependency
+                            self._openai_client = openai.OpenAI(api_key=api_key)  # type: ignore[attr-defined]
+                        except Exception:  # pragma: no cover - if instantiation fails
+                            self._openai_client = None
+                if self._openai_client is None and not hasattr(openai, "ChatCompletion"):
+                    log_event("openai_feedback_unavailable", reason="openai_library_missing")
+                    self.provider = "fallback"
+            if self.provider == "local":
+                self.model = os.getenv("COMPAIR_LOCAL_GENERATION_MODEL", "local-feedback")
+                base_url = os.getenv("COMPAIR_LOCAL_MODEL_URL", "http://local-model:9000")
+                route = os.getenv("COMPAIR_LOCAL_GENERATION_ROUTE", "/generate")
+                self.endpoint = f"{base_url.rstrip('/')}{route}"
+            else:
+                self.model = "external"
+                self.endpoint = None
 
     @property
     def is_cloud(self) -> bool:
         return self._cloud_impl is not None
 
 
+def _reference_snippets(references: Iterable[Any], limit: int = 3) -> List[str]:
+    snippets: List[str] = []
+    for ref in references:
+        snippet = getattr(ref, "content", "") or ""
+        snippet = snippet.replace("\n", " ").strip()
+        if snippet:
+            snippets.append(snippet[:200])
+        if len(snippets) == limit:
+            break
+    return snippets
+
+
 def _fallback_feedback(text: str, references: list[Any]) -> str:
-    if not references:
-        return "NONE"
-    top_ref = references[0]
-    snippet = getattr(top_ref, "content", "") or ""
-    snippet = snippet.replace("\n", " ").strip()[:200]
-    if not snippet:
+    snippets = _reference_snippets(references)
+    if not snippets:
         return "NONE"
-    return f"Check alignment with this reference: {snippet}"
+    joined = "; ".join(snippets)
+    return f"Consider aligning with these reference passages: {joined}"
 
 
-def get_feedback(
+def _openai_feedback(
     reviewer: Reviewer,
     doc: Document,
     text: str,
     references: list[Any],
     user: User,
-) -> str:
-    if reviewer.is_cloud and cloud_get_feedback is not None:
-        return cloud_get_feedback(reviewer._cloud_impl, doc, text, references, user)  # type: ignore[arg-type]
+) -> str | None:
+    if openai is None:
+        return None
+    instruction = reviewer.length_map.get(user.preferred_feedback_length, "1–2 short sentences")
+    ref_text = "\n\n".join(_reference_snippets(references, limit=3))
+    messages = [
+        {
+            "role": "system",
+            "content": (
+                "You are Compair, an assistant that delivers concise, actionable feedback on a user's document. "
+                "Focus on clarity, cohesion, and usefulness."
+            ),
+        },
+        {
+            "role": "user",
+            "content": (
+                f"Document:\n{text}\n\nHelpful reference excerpts:\n{ref_text or 'None provided'}\n\n"
+                f"Respond with {instruction} that highlights the most valuable revision to make next."
+            ),
+        },
+    ]
+
+    try:
+        if reviewer._openai_client is not None and hasattr(reviewer._openai_client, "responses"):
+            response = reviewer._openai_client.responses.create(  # type: ignore[union-attr]
+                model=reviewer.openai_model,
+                input=messages,
+                max_output_tokens=256,
+            )
+            content = getattr(response, "output_text", None)
+            if not content and hasattr(response, "outputs"):
+                # Legacy compatibility: join content parts
+                parts = []
+                for item in getattr(response, "outputs", []):
+                    parts.extend(getattr(item, "content", []))
+                content = " ".join(getattr(part, "text", "") for part in parts)
+        elif hasattr(openai, "ChatCompletion"):
+            chat_response = openai.ChatCompletion.create(  # type: ignore[attr-defined]
+                model=reviewer.openai_model,
+                messages=messages,
+                temperature=0.3,
+                max_tokens=256,
+            )
+            content = (
+                chat_response["choices"][0]["message"]["content"].strip()  # type: ignore[index, assignment]
+            )
+        else:
+            content = None
+    except Exception as exc:  # pragma: no cover - network/API failure
+        log_event("openai_feedback_failed", error=str(exc))
+        content = None
+    if content:
+        content = content.strip()
+        if content:
+            return content
+    return None
 
+
+def _local_feedback(
+    reviewer: Reviewer,
+    text: str,
+    references: list[Any],
+    user: User,
+) -> str | None:
     payload = {
         "document": text,
         "references": [getattr(ref, "content", "") for ref in references],
-        "length_instruction": {
-            "Brief": "1–2 short sentences",
-            "Detailed": "A couple short paragraphs",
-            "Verbose": "As thorough as reasonably possible without repeating information",
-        }.get(user.preferred_feedback_length, "1–2 short sentences"),
+        "length_instruction": reviewer.length_map.get(
+            user.preferred_feedback_length,
+            "1–2 short sentences",
+        ),
     }
 
     try:
         response = requests.post(reviewer.endpoint, json=payload, timeout=30)
         response.raise_for_status()
         data = response.json()
-        feedback = data.get("feedback")
+        feedback = data.get("feedback") or data.get("text")
         if feedback:
-            return feedback
+            return str(feedback).strip()
     except Exception as exc:  # pragma: no cover - network failures stay graceful
         log_event("local_feedback_failed", error=str(exc))
 
+    return None
+
+
+def get_feedback(
+    reviewer: Reviewer,
+    doc: Document,
+    text: str,
+    references: list[Any],
+    user: User,
+) -> str:
+    if reviewer.is_cloud and cloud_get_feedback is not None:
+        return cloud_get_feedback(reviewer._cloud_impl, doc, text, references, user)  # type: ignore[arg-type]
+
+    if reviewer.provider == "openai":
+        feedback = _openai_feedback(reviewer, doc, text, references, user)
+        if feedback:
+            return feedback
+
+    if reviewer.provider == "local" and getattr(reviewer, "endpoint", None):
+        feedback = _local_feedback(reviewer, text, references, user)
+        if feedback:
+            return feedback
+
     return _fallback_feedback(text, references)

compair_core/compair/models.py

@@ -76,9 +76,13 @@ def _embedding_column():
             raise RuntimeError(
                 "pgvector is required when COMPAIR_VECTOR_BACKEND is set to 'pgvector'."
             )
-        return mapped_column(Vector(EMBEDDING_DIMENSION), nullable=True)
+        return mapped_column(
+            Vector(EMBEDDING_DIMENSION),
+            nullable=True,
+            default=None,
+        )
     # Store embeddings as JSON arrays (works across SQLite/Postgres without pgvector)
-    return mapped_column(JSON, nullable=True)
+    return mapped_column(JSON, nullable=True, default=None)
 
 
 def cosine_similarity(vec1: Sequence[float] | None, vec2: Sequence[float] | None) -> float | None:
@@ -279,10 +283,10 @@ class Document(BaseObject):
     doc_type: Mapped[str]
     datetime_created: Mapped[datetime]
     datetime_modified: Mapped[datetime]
+    embedding: Mapped[list[float] | None] = _embedding_column()
     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[list[float] | None] = _embedding_column()
 
     user = relationship("User", back_populates="documents")
     groups = relationship("Group", secondary="document_to_group", back_populates="documents")
@@ -315,8 +319,8 @@ class Note(Base):
     author_id: Mapped[str] = mapped_column(ForeignKey("user.user_id", ondelete="CASCADE"), index=True)
     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[list[float] | None] = _embedding_column()
+    datetime_created: Mapped[datetime] = mapped_column(default=datetime.now(timezone.utc))
 
     document = relationship("Document", back_populates="notes")
     author = relationship("User", back_populates="notes")

compair_core/server/local_model/app.py

@@ -46,13 +46,19 @@ class EmbedResponse(BaseModel):
 
 
 class GenerateRequest(BaseModel):
+    # Legacy format used by the CLI shim
     system: str | None = None
-    prompt: str
+    prompt: str | None = None
     verbosity: str | None = None
 
+    # Core API payload (document + references)
+    document: str | None = None
+    references: List[str] | None = None
+    length_instruction: str | None = None
+
 
 class GenerateResponse(BaseModel):
-    text: str
+    feedback: str
 
 
 @app.post("/embed", response_model=EmbedResponse)
@@ -62,12 +68,20 @@ def embed(request: EmbedRequest) -> EmbedResponse:
 
 @app.post("/generate", response_model=GenerateResponse)
 def generate(request: GenerateRequest) -> GenerateResponse:
-    prompt = request.prompt.strip()
-    if not prompt:
-        return GenerateResponse(text="NONE")
-
-    first_sentence = prompt.split("\n", 1)[0][:200]
-    verbosity = request.verbosity or "default"
-    return GenerateResponse(
-        text=f"[local-{verbosity}] Key takeaway: {first_sentence}"
-    )
+    # Determine the main text input (document or prompt)
+    text_input = request.document or request.prompt or ""
+    text_input = text_input.strip()
+
+    if not text_input:
+        return GenerateResponse(feedback="NONE")
+
+    first_sentence = text_input.split("\n", 1)[0][:200]
+    verbosity = request.length_instruction or request.verbosity or "brief response"
+    ref_snippet = ""
+    if request.references:
+        top_ref = (request.references[0] or "").strip()
+        if top_ref:
+            ref_snippet = f" Reference: {top_ref[:160]}"
+
+    feedback = f"[local-feedback] {verbosity}: {first_sentence}{ref_snippet}".strip()
+    return GenerateResponse(feedback=feedback or "NONE")

{compair_core-0.3.14.dist-info → compair_core-0.3.15.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compair-core
-Version: 0.3.14
+Version: 0.3.15
 Summary: Open-source foundation of the Compair collaboration platform.
 Author: RocketResearch, Inc.
 License: MIT
@@ -94,6 +94,8 @@ Key environment variables for the core edition:
 - `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).
+- `COMPAIR_GENERATION_PROVIDER` (`local`) – choose how feedback is produced. Options: `local` (call the bundled FastAPI service), `openai` (use ChatGPT-compatible APIs with an API key), or `fallback` (skip generation and surface similar references only).
+- `COMPAIR_OPENAI_API_KEY` / `COMPAIR_OPENAI_MODEL` – when using the OpenAI provider, supply your API key and optional model name (defaults to `gpt-4o-mini`). The fallback kicks in automatically if the key or SDK is unavailable.
 
 See `compair_core/server/settings.py` for the full settings surface.
 

{compair_core-0.3.14.dist-info → compair_core-0.3.15.dist-info}/RECORD

@@ -4,10 +4,10 @@ compair_core/compair/__init__.py,sha256=V2mqe6UQEvY4U8XL8T-TtCRNDWVUMNeCLZ8nsYQL
 compair_core/compair/celery_app.py,sha256=OM_Saza9yC9Q0kz_WXctfswrKkG7ruT52Zl5E4guiT0,640
 compair_core/compair/default_groups.py,sha256=dbacrFkSjqEQZ_uoFU5gYhgIoP_3lmvz6LJNHCJvxlw,498
 compair_core/compair/embeddings.py,sha256=hlDt_C4JWKQCx_pWZZeBSNaF2gvkIUupQyl5Wq4HR34,2754
-compair_core/compair/feedback.py,sha256=jgDxYKo5PzW2p-uLf5ETmlL-mDAfqzJazO1NZDK-Z-g,2755
+compair_core/compair/feedback.py,sha256=FvHefla0bKdyWLNduAVjCxASgZAM-XumElht3JtWHPs,7573
 compair_core/compair/logger.py,sha256=mB9gV3FfC0qE_G9NBErpEAJhyGxDxEVKqWKu3n8hOkc,802
 compair_core/compair/main.py,sha256=d2d7ENXDWKVG_anOG8Kkr-r9W19h__7zwXSo-3zx4rY,8976
-compair_core/compair/models.py,sha256=Bepyo7Vyh9serzVsE0-s0HXo1oKDNem6kV-DnqyeFwU,16515
+compair_core/compair/models.py,sha256=0ysM2HG1E8j4MmW6P7abEgNFfor_yfFffScYIl8xku8,16590
 compair_core/compair/schema.py,sha256=TxQpDQ96J_tIj-Y_1C_x2eUYn9n_LOG6XiNLCX1-GYY,2902
 compair_core/compair/tasks.py,sha256=CY-Oo9UeqNF52s2twUkXSyHGgz1_PX8ffTqxmmIlDJs,3864
 compair_core/compair/utils.py,sha256=7bbkRb-9zmPAogqR7zQHLnQUkoeW9bwhqPz2S5FRKSc,978
@@ -21,7 +21,7 @@ compair_core/server/app.py,sha256=4FY4Pur1fJUkQAwBApLhPQsnF6BRNZjUQ-X7bJCv2mI,36
 compair_core/server/deps.py,sha256=0X-Z5JQGeXwbMooWIOC2kXVmsiJIvgUtqkK2PmDjKpI,1557
 compair_core/server/settings.py,sha256=mWE5vgIx3jxm6LzyeH6QL1tCwxQ6bsZIQffVowqtkFQ,1681
 compair_core/server/local_model/__init__.py,sha256=YlzDgorgAjGou9d_W29Xp3TVu08e4t9x8csFxn8cgSE,50
-compair_core/server/local_model/app.py,sha256=N0cb1BdBQbp2TIwMeh4O4ttdgF7k-uHoURnI_X8HPFE,1943
+compair_core/server/local_model/app.py,sha256=943dnixGcaxHixBXVgJmG5G69gisFcwcIh65NJ5scGU,2552
 compair_core/server/providers/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 compair_core/server/providers/console_mailer.py,sha256=7ady954yOPlT8t3spkDvdMdO3BTwosJUj1cpVNOwj8U,308
 compair_core/server/providers/contracts.py,sha256=pYA_2AaPHw089O_UP2JtWRHbIiVkoNGhLuesuNATowU,1858
@@ -31,8 +31,8 @@ compair_core/server/providers/noop_billing.py,sha256=V18Cpl1D1reM3xhgw-lShGliVpY
 compair_core/server/providers/noop_ocr.py,sha256=fMaJrivDef38-ECgIuTXUBCIm_avgvZf3nQ3UTdFPNI,341
 compair_core/server/routers/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 compair_core/server/routers/capabilities.py,sha256=2U9lEzyQRRYftprrvEeM5Lif_5rhiRGqZhIsvYZBaE4,1349
-compair_core-0.3.14.dist-info/licenses/LICENSE,sha256=OXLcl0T2SZ8Pmy2_dmlvKuetivmyPd5m1q-Gyd-zaYY,35149
-compair_core-0.3.14.dist-info/METADATA,sha256=V6ofQ2FtQI3SnhcnseQeP3DNx_HiUNx-8yD4vzpTQvA,5703
-compair_core-0.3.14.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-compair_core-0.3.14.dist-info/top_level.txt,sha256=1dpwoLSY2DWQUVGS05Tq0MuFXg8sabYzg4V2deLzzuo,13
-compair_core-0.3.14.dist-info/RECORD,,
+compair_core-0.3.15.dist-info/licenses/LICENSE,sha256=OXLcl0T2SZ8Pmy2_dmlvKuetivmyPd5m1q-Gyd-zaYY,35149
+compair_core-0.3.15.dist-info/METADATA,sha256=GPBU-Wg4GLnY3isvgl4z9JTtTQs-iKyneD_rTMNB7hg,6191
+compair_core-0.3.15.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+compair_core-0.3.15.dist-info/top_level.txt,sha256=1dpwoLSY2DWQUVGS05Tq0MuFXg8sabYzg4V2deLzzuo,13
+compair_core-0.3.15.dist-info/RECORD,,