Improve RAG recall: more chunks, synonym expansion

Author: medevs

backend/app/agent/core.py

@@ -74,8 +74,8 @@ async def run(self, query: str, history: Optional[List[Dict[str, str]]] = None,
         logger.info(f"[RAG] Query: {query[:50]}...")
 
         # Step 1: Execute BOTH searches (hybrid)
-        semantic_results = await self.semantic_tool.execute(query=query, limit=5)
-        keyword_results = await self.keyword_tool.execute(query=query, limit=5)
+        semantic_results = await self.semantic_tool.execute(query=query, limit=10)
+        keyword_results = await self.keyword_tool.execute(query=query, limit=10)
 
         # Step 2: Merge with RRF and log which path was used
         if semantic_results and keyword_results:
@@ -97,9 +97,9 @@ async def run(self, query: str, history: Optional[List[Dict[str, str]]] = None,
             preview = doc.get('content', '')[:80].replace('\n', ' ')
             logger.info(f"Doc {i+1}: {source} - {preview}...")
 
-        # Step 3: Build context and generate answer
+        # Step 3: Build context and generate answer (use top 8 chunks for better coverage)
         context_parts = []
-        for doc in merged_docs[:5]:
+        for doc in merged_docs[:8]:
             source = doc.get('metadata', {}).get('source', 'unknown')
             content = doc.get('content', '')
             if content:

backend/app/services/tools/keyword_search.py

@@ -13,6 +13,20 @@
               'at', 'to', 'for', 'of', 'with', 'by', 'about', 'can', 'could', 'would',
               'should', 'have', 'has', 'had', 'be', 'been', 'being', 'this', 'that'}
 
+# Query expansion: synonyms for better recall
+SYNONYMS = {
+    'job': ['work', 'position', 'role', 'employment', 'experience'],
+    'work': ['job', 'position', 'role', 'employment', 'experience'],
+    'company': ['employer', 'organization', 'firm'],
+    'skill': ['skills', 'technology', 'technologies', 'expertise'],
+    'skills': ['skill', 'technology', 'technologies', 'expertise'],
+    'education': ['degree', 'diploma', 'university', 'school', 'training'],
+    'project': ['projects', 'portfolio', 'work'],
+    'projects': ['project', 'portfolio', 'work'],
+    'language': ['languages', 'speak', 'fluent'],
+    'languages': ['language', 'speak', 'fluent'],
+}
+
 
 class KeywordSearchTool(BaseTool):
     name = "keyword_search"
@@ -21,7 +35,7 @@ class KeywordSearchTool(BaseTool):
     def _build_tsquery(self, query: str) -> str:
         """
         Build OR-based tsquery from natural language query.
-        Filters stop words and joins remaining terms with OR (|).
+        Filters stop words, expands synonyms, joins with OR (|).
         """
         # Extract words, lowercase, filter stop words
         words = re.findall(r'\b\w+\b', query.lower())
@@ -31,8 +45,14 @@ def _build_tsquery(self, query: str) -> str:
             # Fallback: use all words if filtering removed everything
             keywords = [w for w in words if len(w) > 2]
 
+        # Expand with synonyms for better recall
+        expanded = set(keywords)
+        for word in keywords:
+            if word in SYNONYMS:
+                expanded.update(SYNONYMS[word])
+
         # Join with OR operator for PostgreSQL tsquery
-        return ' | '.join(keywords) if keywords else query
+        return ' | '.join(expanded) if expanded else query
 
     async def execute(self, query: str, limit: int = 10) -> List[Dict[str, Any]]:
         """