opensolr-haystack 0.2.5__tar.gz → 0.2.7__tar.gz

{opensolr_haystack-0.2.5 → opensolr_haystack-0.2.7}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: opensolr-haystack
-Version: 0.2.5
+Version: 0.2.7
 Summary: Haystack integration for Opensolr — managed Apache Solr DocumentStore with server-side embeddings and hybrid BM25+kNN retrieval
 Author-email: Opensolr <support@opensolr.com>
 License: MIT
@@ -22,6 +22,8 @@ Dynamic: license-file
 [Opensolr](https://opensolr.com) — managed Apache Solr as a DocumentStore,
 with **server-side embeddings** and native **hybrid (BM25 + kNN) retrieval**.
 
+**See it live (real news index, hybrid + AI answer):** https://search.opensolr.com/news__dense?q=how+am+I+supposed+to+save+money%3F
+
 No embedder components needed in your pipeline — texts and queries are
 embedded on Opensolr's GPU infrastructure (multilingual E5-large-instruct,
 1024 dimensions, cosine).

{opensolr_haystack-0.2.5 → opensolr_haystack-0.2.7}/README.md

@@ -4,6 +4,8 @@
 [Opensolr](https://opensolr.com) — managed Apache Solr as a DocumentStore,
 with **server-side embeddings** and native **hybrid (BM25 + kNN) retrieval**.
 
+**See it live (real news index, hybrid + AI answer):** https://search.opensolr.com/news__dense?q=how+am+I+supposed+to+save+money%3F
+
 No embedder components needed in your pipeline — texts and queries are
 embedded on Opensolr's GPU infrastructure (multilingual E5-large-instruct,
 1024 dimensions, cosine).

{opensolr_haystack-0.2.5 → opensolr_haystack-0.2.7}/haystack_integrations/document_stores/opensolr/client.py

@@ -42,6 +42,26 @@ def resolve_location(location: str) -> str:
 BATCH_EMBED_MAX = 50
 
 
+# Default RAG instruction (added 2026-08-25).
+# Without an instruction the server falls back to "Answer the query based on
+# the context", which makes the model hedge: it opens with a disclaimer that
+# the context does not cover the question and then answers it anyway. These
+# rules make it lead with the answer and keep the concrete details. Callers
+# can still override everything by passing ``instruction``.
+DEFAULT_RAG_INSTRUCTION = (
+    # Consolidated 2026-08-26. The rule list had grown to eight overlapping lines and the
+    # 3B model drowned in them: it restated the question, opened with a preamble, and
+    # presented items as fit for purposes the context never stated. Tested against the
+    # live model on three cases (answer present, answer absent, yes/no) before shipping.
+    "Answer the question using only the context below.\n"
+    "Begin with the answer itself, the specific fact, product or detail. "
+    "No preamble, no restating the question, no heading.\n"
+    "Then give the concrete supporting details from the context: names, numbers, sizes, dates. "
+    "Write two to four sentences.\n"
+    "If the context holds no answer, say that in the first sentence and name what it does contain instead.\n"
+    "Never present something as suitable for a purpose the context does not state."
+)
+
 class OpensolrError(RuntimeError):
     """Raised when an Opensolr API call fails."""
 
@@ -353,6 +373,16 @@ class OpensolrClient:
                 index, query, rows=docs, fl="title,description,text", fq=fq
             )
             hits = body.get("response", {}).get("docs", [])
+        # Relevance floor (2026-08-25). Retrieval always returns `docs` hits, so a
+        # narrow question arrives with one good match and several unrelated ones.
+        # The model then sees that most of its context does not answer the query
+        # and hedges. Drop anything scoring below half of the best hit; documents
+        # without a score are kept, since only provable weakness is filtered.
+        scored = [float(h["score"]) for h in hits if isinstance(h, dict) and h.get("score") is not None]
+        if scored:
+            floor = max(scored) * 0.5
+            hits = [h for h in hits if h.get("score") is None or float(h["score"]) >= floor]
+
         parts: List[str] = []
         for doc in hits[:docs]:
             text_words = _flat(doc.get("text")).split()[:words]
@@ -390,8 +420,7 @@ class OpensolrClient:
             "stream": "false",
             **params,
         }
-        if instruction:
-            data["instruction"] = instruction
+        data["instruction"] = instruction or DEFAULT_RAG_INSTRUCTION
         if "context" not in data:
             try:
                 context = self._rag_context(

{opensolr_haystack-0.2.5 → opensolr_haystack-0.2.7}/opensolr_haystack.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: opensolr-haystack
-Version: 0.2.5
+Version: 0.2.7
 Summary: Haystack integration for Opensolr — managed Apache Solr DocumentStore with server-side embeddings and hybrid BM25+kNN retrieval
 Author-email: Opensolr <support@opensolr.com>
 License: MIT
@@ -22,6 +22,8 @@ Dynamic: license-file
 [Opensolr](https://opensolr.com) — managed Apache Solr as a DocumentStore,
 with **server-side embeddings** and native **hybrid (BM25 + kNN) retrieval**.
 
+**See it live (real news index, hybrid + AI answer):** https://search.opensolr.com/news__dense?q=how+am+I+supposed+to+save+money%3F
+
 No embedder components needed in your pipeline — texts and queries are
 embedded on Opensolr's GPU infrastructure (multilingual E5-large-instruct,
 1024 dimensions, cosine).

{opensolr_haystack-0.2.5 → opensolr_haystack-0.2.7}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "opensolr-haystack"
-version = "0.2.5"
+version = "0.2.7"
 description = "Haystack integration for Opensolr — managed Apache Solr DocumentStore with server-side embeddings and hybrid BM25+kNN retrieval"
 readme = "README.md"
 license = { text = "MIT" }