Add single-pass CIM processor: 2 LLM calls, ~2.5 min processing

New processing strategy `single_pass_quality_check` replaces the multi-pass
agentic RAG pipeline (15-25 min) with a streamlined 2-call approach:

1. Full-document LLM extraction (Sonnet) — single call with complete CIM text
2. Delta quality-check (Haiku) — reviews extraction, returns only corrections

Key changes:
- New singlePassProcessor.ts with extraction + quality check flow
- llmService: qualityCheckCIMDocument() with delta-only corrections array
- llmService: improved prompt requiring professional inferences for qualitative
  fields instead of defaulting to "Not specified in CIM"
- Removed deterministic financial parser from single-pass flow (LLM outperforms
  it — parser matched footnotes and narrative text as financials)
- Default strategy changed to single_pass_quality_check
- Completeness scoring with diagnostic logging of empty fields

Tested on 2 real CIMs: 100% completeness, correct financials, ~150s each.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

backend/sql/fix_vector_search_timeout.sql

@@ -1,12 +1,12 @@
--- Fix vector search timeout by adding document_id filtering and optimizing the query
--- This prevents searching across all documents and only searches within a specific document
+-- Fix vector search timeout by pre-filtering on document_id BEFORE vector search
+-- When document_id is provided, this avoids the full IVFFlat index scan (26K+ rows)
+-- and instead computes distances on only ~80 chunks per document.
 
--- Drop the old function (handle all possible signatures)
+-- Drop old function signatures
 DROP FUNCTION IF EXISTS match_document_chunks(vector(1536), float, int);
 DROP FUNCTION IF EXISTS match_document_chunks(vector(1536), float, int, text);
 
--- Create optimized function with document_id filtering
--- document_id is TEXT (varchar) in the actual schema
+-- Create optimized function that branches based on whether document_id is provided
 CREATE OR REPLACE FUNCTION match_document_chunks (
   query_embedding vector(1536),
   match_threshold float,
@@ -15,29 +15,51 @@ CREATE OR REPLACE FUNCTION match_document_chunks (
 )
 RETURNS TABLE (
   id UUID,
-  document_id TEXT,
+  document_id VARCHAR(255),
   content text,
   metadata JSONB,
   chunk_index INT,
   similarity float
 )
-LANGUAGE sql STABLE
+LANGUAGE plpgsql STABLE
 AS $$
-  SELECT
-    document_chunks.id,
-    document_chunks.document_id,
-    document_chunks.content,
-    document_chunks.metadata,
-    document_chunks.chunk_index,
-    1 - (document_chunks.embedding <=> query_embedding) AS similarity
-  FROM document_chunks
-  WHERE document_chunks.embedding IS NOT NULL
-    AND (filter_document_id IS NULL OR document_chunks.document_id = filter_document_id)
-    AND 1 - (document_chunks.embedding <=> query_embedding) > match_threshold
-  ORDER BY document_chunks.embedding <=> query_embedding
-  LIMIT match_count;
+BEGIN
+  IF filter_document_id IS NOT NULL THEN
+    -- FAST PATH: Pre-filter by document_id using btree index, then compute
+    -- vector distances on only that document's chunks (~80 rows).
+    -- This completely bypasses the IVFFlat index scan.
+    RETURN QUERY
+    SELECT
+      dc.id,
+      dc.document_id,
+      dc.content,
+      dc.metadata,
+      dc.chunk_index,
+      1 - (dc.embedding <=> query_embedding) AS similarity
+    FROM document_chunks dc
+    WHERE dc.document_id = filter_document_id
+      AND dc.embedding IS NOT NULL
+      AND 1 - (dc.embedding <=> query_embedding) > match_threshold
+    ORDER BY dc.embedding <=> query_embedding
+    LIMIT match_count;
+  ELSE
+    -- SLOW PATH: Search across all documents using IVFFlat index.
+    -- Only used when no document_id filter is provided.
+    RETURN QUERY
+    SELECT
+      dc.id,
+      dc.document_id,
+      dc.content,
+      dc.metadata,
+      dc.chunk_index,
+      1 - (dc.embedding <=> query_embedding) AS similarity
+    FROM document_chunks dc
+    WHERE dc.embedding IS NOT NULL
+      AND 1 - (dc.embedding <=> query_embedding) > match_threshold
+    ORDER BY dc.embedding <=> query_embedding
+    LIMIT match_count;
+  END IF;
+END;
 $$;
 
--- Add comment explaining the optimization
-COMMENT ON FUNCTION match_document_chunks IS 'Optimized vector search that filters by document_id first to prevent timeouts. Always pass filter_document_id when searching within a specific document.';
-
+COMMENT ON FUNCTION match_document_chunks IS 'Vector search with fast document-scoped path. When filter_document_id is provided, uses btree index to pre-filter (~80 rows) instead of scanning the full IVFFlat index (26K+ rows).';