diff --git a/nlp-service/requirements.txt b/nlp-service/requirements.txt
index 19491d0..44435b2 100644
--- a/nlp-service/requirements.txt
+++ b/nlp-service/requirements.txt
@@ -1,21 +1,21 @@
 # Kompl v2 nlp-service — commit 4. Pins per docs/research/2026-04-08-conversion-deps.md
 # and docs/research/2026-04-09-llm-compile.md.
-fastapi==0.136.1
-uvicorn[standard]==0.47.0
+fastapi==0.136.3
+uvicorn[standard]==0.49.0
 pydantic==2.13.4
 httpx==0.28.1
-markitdown[all]==0.1.5
+markitdown[all]==0.1.6
 # Promoted to a direct dep so routers/conversion.py can call
 # youtube-transcript-api WITHOUT going through MarkItDown for YouTube URLs.
 # Reason: MarkItDown's YouTubeConverter silently falls back to scraping the
 # watch-page HTML when no transcript exists (returns ~800 chars of footer
 # chrome — verified live on session 4a00f339). We need the transcript-vs-no-
 # transcript signal explicitly, which only the direct API exposes.
-# markitdown[all]==0.1.5 constrains ~=1.0.0; Dockerfile upgrades to 1.2.x
+# markitdown[all]==0.1.6 constrains ~=1.0.0; Dockerfile upgrades to 1.2.x
 # after pip install (1.0.x breaks on current YouTube timedtext responses).
 youtube-transcript-api~=1.0.0
 # Commit 4: LLM compile. Use google-genai NOT google-generativeai (deprecated 2025-Q1).
-google-genai==2.4.0
+google-genai==2.8.0
 # Commit 4: async token-bucket rate limiter. Single uvicorn worker only —
 # InMemoryBucket is process-local; see research artifact section 3.
 pyrate-limiter==3.7.0
@@ -24,7 +24,7 @@ spacy==3.8.14
 rake-nltk==1.0.6
 yake==0.7.3
 keybert==0.9.0
-sentence-transformers==5.5.0
+sentence-transformers==5.5.1
 # Dockerfile pins torch==2.5.1 (meta-tensor compat). transformers 5.x imports
 # torch.float8_e8m0fnu at load time, which 2.5.1 lacks — breaks
 # /resolve/embedding on fresh pip installs. Cap at 4.x; pytest stubs ST so
@@ -32,7 +32,7 @@ sentence-transformers==5.5.0
 transformers>=4.41.0,<5.0.0
 # pytextrank: spaCy-native TextRank component (PyPI-available, no git dep).
 pytextrank==3.3.0
-scikit-learn==1.8.0
+scikit-learn==1.9.0
 # nltk: required by rake-nltk (stopwords + punkt tokenizer data).
 nltk==3.9.4
 # Part 2b: fast Levenshtein/Jaro-Winkler matching for entity resolution.
@@ -42,7 +42,7 @@ rapidfuzz>=3.14.5
 json-repair>=0.59.10
 # Commit 7: Chroma vector store (embedded, no separate server). 0.4.x API:
 # PersistentClient, get_or_create_collection, cosine distance.
-chromadb==0.4.24
+chromadb==0.6.3
 # chromadb 0.4.24 references np.float_ (chromadb/api/types.py:102), which
 # NumPy 2.0 removed. Pin numpy<2 until chromadb is upgraded.
 numpy<2