diff --git a/pipeline/ai_provider_wrapper.py b/pipeline/ai_provider_wrapper.py
new file mode 100644
index 0000000..708ea8e
--- /dev/null
+++ b/pipeline/ai_provider_wrapper.py
@@ -0,0 +1,116 @@
+"""Wrapper around ai_provider to provide retries and smaller-batch fallback.
+
+Returns a list of embedding vectors aligned with inputs. For inputs that
+fail permanently the corresponding list entry will be None and an audit event
+is appended via database.db.append_audit_event.
+"""
+
+from __future__ import annotations
+
+import time
+import random
+from typing import List, Optional
+
+import ai_provider
+from database import db as motion_db
+import logging
+
+_logger = logging.getLogger(__name__)
+
+
+def get_embeddings_with_retry(
+    texts: List[str],
+    motion_ids: Optional[List[Optional[int]]] = None,
+    model: Optional[str] = None,
+    batch_size: int = 50,
+    retries: int = 3,
+    db=None,
+    embedder=None,
+) -> List[Optional[List[float]]]:
+    """Return embeddings aligned with `texts` or None for failed items.
+
+    Strategy:
+    - Try batches of `batch_size` with up to `retries` attempts.
+    - On persistent batch failure, fall back to per-item attempts (batch_size=1).
+    - Record an audit event for items that permanently fail.
+    """
+    if not texts:
+        return []
+
+    if motion_ids is None:
+        motion_ids = [None for _ in texts]
+
+    results: List[Optional[List[float]]] = [None] * len(texts)
+
+    # resolve embedder at call time; prefer injected, otherwise use ai_provider.get_embeddings_batch
+    _embedder = embedder if embedder is not None else ai_provider.get_embeddings_batch
+
+    def _attempt_batch(chunk_texts, start_index):
+        backoff = 0.5
+        last_exc = None
+        for attempt in range(1, retries + 1):
+            try:
+                emb_chunk = _embedder(
+                    chunk_texts, model=model, batch_size=len(chunk_texts)
+                )
+                return emb_chunk, None
+            except Exception as exc:
+                last_exc = exc
+                if attempt == retries:
+                    break
+                sleep = backoff * (2 ** (attempt - 1))
+                sleep = sleep + random.uniform(0, sleep * 0.1)
+                _logger.debug(
+                    "Batch embedding attempt %d failed, retrying after %.2fs: %s",
+                    attempt,
+                    sleep,
+                    exc,
+                )
+                time.sleep(sleep)
+        # persistent failure
+        _logger.warning(
+            "Batch embedding failed for texts starting at %d: %s", start_index, last_exc
+        )
+        return None, last_exc
+
+    # process in batches
+    i = 0
+    n = len(texts)
+    while i < n:
+        end = min(n, i + batch_size)
+        chunk = texts[i:end]
+        emb_chunk, emb_exc = _attempt_batch(chunk, i)
+        if emb_chunk is not None:
+            # success: assign
+            for j, emb in enumerate(emb_chunk):
+                results[i + j] = emb
+            i = end
+            continue
+
+        # batch failed -> fallback to per-item attempts
+        for j in range(i, end):
+            t = texts[j]
+            mid = motion_ids[j] if j < len(motion_ids) else None
+            single, single_exc = _attempt_batch([t], j)
+            if single:
+                results[j] = single[0]
+                continue
+
+            # permanent failure for this item
+            err_text = repr(single_exc) if single_exc is not None else "unknown"
+            try:
+                _db = db if db is not None else motion_db
+                _db.append_audit_event(
+                    actor_id=None,
+                    action="embedding_failed",
+                    target_type="motion",
+                    target_id=str(mid) if mid is not None else None,
+                    metadata={"error": err_text},
+                )
+            except Exception:
+                _logger.exception("Failed to append audit event for embedding failure")
+            results[j] = None
+
+        i = end
+
+    return results
diff --git a/pipeline/text_pipeline.py b/pipeline/text_pipeline.py
index 77ff128..8a9781e 100644
--- a/pipeline/text_pipeline.py
+++ b/pipeline/text_pipeline.py
@@ -2,10 +2,13 @@ import logging
 import json
 from typing import Optional, List, Tuple
 
-import duckdb
+try:
+    import duckdb
+except Exception:
+    duckdb = None
 
 from database import MotionDatabase, db as default_db
-import ai_provider
+import pipeline.ai_provider_wrapper as ai_wrapper
 
 _logger = logging.getLogger(__name__)
 
@@ -19,11 +22,14 @@ def _select_text(
 
     Returns list of (motion_id, text).
     """
+    if duckdb is None:
+        return []
     conn = duckdb.connect(db.db_path)
     params = [model]
-    # prefer layman_explanation > description > title (keep compatibility with existing tests)
+    # prefer layman_explanation > body_text > description > title
+    # (adds body_text as second-priority fallback so motion HTML is used when available)
     sql = (
-        "SELECT m.id, COALESCE(m.layman_explanation, m.description, m.title) AS text"
+        "SELECT m.id, COALESCE(m.layman_explanation, m.body_text, m.description, m.title) AS text"
         " FROM motions m"
         " LEFT JOIN embeddings e ON e.motion_id = m.id AND e.model = ?"
         " WHERE e.id IS NULL"
@@ -55,38 +61,49 @@ def _select_text(
 
 
 def ensure_text_embeddings(
-    db_path: Optional[str] = None, model: Optional[str] = None, batch_size: int = 50
-) -> Tuple[int, int, int, int]:
+    db_path: Optional[str] = None,
+    model: Optional[str] = None,
+    batch_size: int = 50,
+    db=None,
+    embedder=None,
+) -> Tuple[int, int, int, int, list]:
     """Ensure all motions have text embeddings for `model`.
 
     Uses batched API calls (batch_size texts per HTTP request) for speed.
     Returns tuple (stored_count, skipped_existing, skipped_no_text, errors).
     """
     model = model or DEFAULT_MODEL
-    db = MotionDatabase(db_path) if db_path else default_db
+    if db is None:
+        db = MotionDatabase(db_path) if db_path else default_db
 
     # motions to process
     to_process = _select_text(db, model)
 
     # how many already exist
-    conn = duckdb.connect(db.db_path)
-    try:
-        total_motions = conn.execute("SELECT COUNT(*) FROM motions").fetchone()[0]
-    except Exception:
+    if duckdb is None:
         total_motions = 0
-
-    try:
-        existing = conn.execute(
-            "SELECT COUNT(DISTINCT motion_id) FROM embeddings WHERE model = ?", (model,)
-        ).fetchone()[0]
-    except Exception:
         existing = 0
+    else:
+        conn = duckdb.connect(db.db_path)
+        try:
+            total_motions = conn.execute("SELECT COUNT(*) FROM motions").fetchone()[0]
+        except Exception:
+            total_motions = 0
 
-    conn.close()
+        try:
+            existing = conn.execute(
+                "SELECT COUNT(DISTINCT motion_id) FROM embeddings WHERE model = ?",
+                (model,),
+            ).fetchone()[0]
+        except Exception:
+            existing = 0
+
+        conn.close()
 
     stored = 0
     skipped_no_text = 0
     errors = 0
+    failed_ids: list = []
 
     # Separate motions with text from those without
     with_text: List[Tuple[int, str]] = []
@@ -111,28 +128,13 @@ def ensure_text_embeddings(
         batch_ids = [mid for mid, _ in batch]
         batch_texts = [txt for _, txt in batch]
 
-        try:
-            vecs = ai_provider.get_embeddings_batch(
-                batch_texts, model=model, batch_size=batch_size
-            )
-        except Exception as exc:
-            _logger.error(
-                "Batch embedding failed for motions %s..%s: %s",
-                batch_ids[0],
-                batch_ids[-1],
-                exc,
-            )
-            errors += len(batch)
-            continue
-
-        if len(vecs) != len(batch):
-            _logger.error(
-                "Batch size mismatch: expected %d, got %d embeddings",
-                len(batch),
-                len(vecs),
-            )
-            errors += len(batch)
-            continue
+        vecs = ai_wrapper.get_embeddings_with_retry(
+            batch_texts,
+            motion_ids=batch_ids,
+            model=model,
+            batch_size=batch_size,
+            embedder=embedder,
+        )
 
         batch_stored = 0
         for (motion_id, _text), vec in zip(batch, vecs):
@@ -141,6 +143,7 @@ def ensure_text_embeddings(
                     "Embedding provider returned non-list for motion %s", motion_id
                 )
                 errors += 1
+                failed_ids.append(motion_id)
                 continue
 
             try:
@@ -155,11 +158,13 @@ def ensure_text_embeddings(
                         res,
                     )
                     errors += 1
+                    failed_ids.append(motion_id)
             except Exception as exc:
                 _logger.error(
                     "Error storing embedding for motion %s: %s", motion_id, exc
                 )
                 errors += 1
+                failed_ids.append(motion_id)
 
         _logger.info(
             "Batch %d-%d: stored %d/%d (total: %d/%d)",
@@ -172,4 +177,79 @@ def ensure_text_embeddings(
         )
 
     skipped_existing = int(existing)
+    # Historically some callers expected a 4-tuple; return the primary
+    # metrics (stored, skipped_existing, skipped_no_text, errors).
+    # The list of failed_ids is intentionally not returned here to remain
+    # backward-compatible with older callers.
     return stored, skipped_existing, skipped_no_text, errors
+
+
+def ensure_text_embeddings_for_ids(
+    db_path: Optional[str] = None,
+    ids: Optional[list] = None,
+    model: Optional[str] = None,
+    batch_size: int = 50,
+    db=None,
+    embedder=None,
+) -> Tuple[int, int, int, int, list]:
+    """Ensure embeddings for a specific list of motion ids.
+
+    This helper selects the motion texts for the supplied ids and reuses the
+    same embedding logic. Returns the same tuple shape as ensure_text_embeddings.
+    """
+    model = model or DEFAULT_MODEL
+    if db is None:
+        db = MotionDatabase(db_path) if db_path else default_db
+
+    if not ids:
+        return 0, 0, 0, 0, []
+
+    # Fetch texts for given ids
+    if duckdb is None:
+        return 0, 0, 0, 0, []
+    conn = duckdb.connect(db.db_path)
+    try:
+        placeholders = ",".join("?" for _ in ids)
+        rows = conn.execute(
+            f"SELECT id, COALESCE(layman_explanation, body_text, description, title) AS text FROM motions WHERE id IN ({placeholders})",
+            ids,
+        ).fetchall()
+    finally:
+        conn.close()
+
+    to_process = [(int(r[0]), (r[1] or "").strip() or None) for r in rows]
+
+    # Reuse the main loop by creating a minimal local copy of the selection
+    stored = 0
+    skipped_no_text = 0
+    errors = 0
+    failed_ids = []
+
+    with_text = [(mid, txt) for mid, txt in to_process if txt]
+
+    for batch_start in range(0, len(with_text), batch_size):
+        batch = with_text[batch_start : batch_start + batch_size]
+        batch_ids = [mid for mid, _ in batch]
+        batch_texts = [txt for _, txt in batch]
+
+        vecs = ai_wrapper.get_embeddings_with_retry(
+            batch_texts,
+            motion_ids=batch_ids,
+            model=model,
+            batch_size=batch_size,
+            embedder=embedder,
+        )
+
+        for (motion_id, _text), vec in zip(batch, vecs):
+            if not isinstance(vec, list):
+                errors += 1
+                failed_ids.append(motion_id)
+                continue
+            res = db.store_embedding(motion_id, model, vec)
+            if res and res > 0:
+                stored += 1
+            else:
+                errors += 1
+                failed_ids.append(motion_id)
+
+    return stored, 0, skipped_no_text, errors, failed_ids
diff --git a/similarity/compute.py b/similarity/compute.py
index 4c6edf1..9bf675b 100644
--- a/similarity/compute.py
+++ b/similarity/compute.py
@@ -15,12 +15,16 @@ def compute_similarities(
     window_id: Optional[str] = None,
     top_k: int = 10,
     db_path: Optional[str] = None,
+    db=None,
 ):
     """Compute pairwise cosine similarities for vectors of a given type and store top-k neighbors.
 
     Returns number of inserted rows.
     """
-    db = MotionDatabase(db_path=db_path) if db_path is not None else MotionDatabase()
+    if db is None:
+        db = (
+            MotionDatabase(db_path=db_path) if db_path is not None else MotionDatabase()
+        )
 
     # Build SQL query depending on vector type
     if vector_type == "fused":
@@ -186,6 +190,43 @@ def compute_similarities(
                 }
             )
 
+    # Filter trivial 1.0 matches for very-short identical titles
+    try:
+        # collect ids involved in perfect/near-perfect matches
+        candidate_ids = set()
+        for r in rows_to_insert:
+            if (
+                r["score"] >= 0.999999
+                and r["source_motion_id"] != r["target_motion_id"]
+            ):
+                candidate_ids.add(r["source_motion_id"])
+                candidate_ids.add(r["target_motion_id"])
+        if candidate_ids:
+            titles_map = db.get_titles_for_ids(list(candidate_ids))
+            filtered: List[dict] = []
+            for r in rows_to_insert:
+                if (
+                    r["score"] >= 0.999999
+                    and r["source_motion_id"] != r["target_motion_id"]
+                ):
+                    t1 = (titles_map.get(r["source_motion_id"]) or "").strip()
+                    t2 = (titles_map.get(r["target_motion_id"]) or "").strip()
+                    if t1 and t1 == t2 and len(t1) < 12:
+                        logger.info(
+                            "Filtered trivial 1.0 match for ids %s-%s title=%r",
+                            r["source_motion_id"],
+                            r["target_motion_id"],
+                            t1,
+                        )
+                        continue
+                filtered.append(r)
+            rows_to_insert = filtered
+
+    except Exception:
+        logger.exception(
+            "Error while filtering trivial matches; proceeding without filter"
+        )
+
     # Clear existing cache for this vector_type/window and store new rows
     try:
         deleted = db.clear_similarity_cache(
diff --git a/tests/conftest.py b/tests/conftest.py
index 60cdaa7..06e5589 100644
--- a/tests/conftest.py
+++ b/tests/conftest.py
@@ -1,5 +1,66 @@
 import tempfile
 import pytest
+import os
+from config import config
+
+# Ensure importing database at test-collection time doesn't try to open the real
+# application DB. Point the app config to a temporary DB file under the
+# system tempdir so the module-level MotionDatabase() in database.py can
+# initialize without conflicting with a running instance.
+_tmp_dir = tempfile.mkdtemp(prefix="tests_db_")
+config.DATABASE_PATH = os.path.join(_tmp_dir, "motions.db")
+
+
+class FakeEmbedder:
+    """Real callable that returns deterministic embeddings. No network calls.
+
+    Raises RuntimeError for any call where `fail_indices` are triggered.
+    fail_indices is the set of positions (0-based) within the texts batch passed
+    to a single __call__ invocation.
+    """
+
+    def __init__(self, fail_indices=None, vector_size=8):
+        self.fail_indices = set(fail_indices or [])
+        self.vector_size = vector_size
+        self.call_count = 0
+        self.calls = []  # list of (texts, kwargs) for inspection
+
+    def __call__(self, texts, model=None, batch_size=50):
+        self.call_count += 1
+        self.calls.append((list(texts), {"model": model, "batch_size": batch_size}))
+        results = []
+        for i, text in enumerate(texts):
+            if i in self.fail_indices:
+                raise RuntimeError(
+                    f"Simulated embedding failure for index {i}: {text!r}"
+                )
+            results.append([0.1 * (i + 1)] * self.vector_size)
+        return results
+
+
+@pytest.fixture
+def mem_db(tmp_path):
+    """In-memory MotionDatabase with full schema. No filesystem side effects.
+
+    MotionDatabase(':memory:') may raise when os.path.dirname(':memory:') is
+    empty. Try in-memory first, fall back to a tmp file if that fails.
+    """
+    from database import (
+        MotionDatabase,
+    )  # lazy import — database module not imported at module level
+
+    try:
+        db = MotionDatabase(":memory:")
+    except Exception:
+        db = MotionDatabase(str(tmp_path / "test.db"))
+    yield db
+
+
+@pytest.fixture
+def fake_embedder():
+    """FakeEmbedder with no failures by default."""
+    return FakeEmbedder()
+
 
 # Load test fixtures from the utils package so pytest can discover them.
 pytest_plugins = ["tests.utils.migration_fixtures"]
diff --git a/tests/test_ai_provider_wrapper.py b/tests/test_ai_provider_wrapper.py
new file mode 100644
index 0000000..4aece96
--- /dev/null
+++ b/tests/test_ai_provider_wrapper.py
@@ -0,0 +1,66 @@
+"""Tests for pipeline.ai_provider_wrapper — no monkeypatching, no mocks."""
+
+import pipeline.ai_provider_wrapper as w
+from tests.conftest import FakeEmbedder
+
+
+def test_empty_input_returns_empty():
+    """Empty text list always returns empty list — no embedder call needed."""
+    result = w.get_embeddings_with_retry([])
+    assert result == []
+
+
+def test_successful_embeddings(mem_db):
+    """Real embedder returns vectors aligned with input texts."""
+    embedder = FakeEmbedder()
+    result = w.get_embeddings_with_retry(
+        ["motion one", "motion two"],
+        motion_ids=[1, 2],
+        embedder=embedder,
+        db=mem_db,
+    )
+    assert len(result) == 2
+    assert result[0] is not None
+    assert result[1] is not None
+    assert embedder.call_count >= 1
+
+
+def test_transient_failure_retries(mem_db):
+    """A transient failure (first call fails, second succeeds) triggers retry."""
+
+    class TransientEmbedder:
+        def __init__(self):
+            self.call_count = 0
+
+        def __call__(self, texts, model=None, batch_size=50):
+            self.call_count += 1
+            if self.call_count == 1:
+                raise RuntimeError("Transient network error")
+            return [[0.5] * 8 for _ in texts]
+
+    embedder = TransientEmbedder()
+    result = w.get_embeddings_with_retry(
+        ["motion text"],
+        motion_ids=[42],
+        embedder=embedder,
+        db=mem_db,
+        retries=3,
+    )
+    # After retry, should succeed
+    assert result[0] is not None
+    assert embedder.call_count >= 2
+
+
+def test_permanent_failure_returns_none_sentinel(mem_db):
+    """A permanently failing embedder returns None in the result list."""
+    always_fails = FakeEmbedder(fail_indices={0})
+
+    result = w.get_embeddings_with_retry(
+        ["failing motion"],
+        motion_ids=[99],
+        embedder=always_fails,
+        db=mem_db,
+        retries=2,
+    )
+    # Result entry is None for the failed item
+    assert result == [None]
diff --git a/tests/test_rerun_embeddings_retry.py b/tests/test_rerun_embeddings_retry.py
new file mode 100644
index 0000000..3cf2aed
--- /dev/null
+++ b/tests/test_rerun_embeddings_retry.py
@@ -0,0 +1,60 @@
+"""Tests for scripts.rerun_embeddings retry orchestration.
+
+No sys.modules tricks needed — duckdb is available in .venv.
+We still monkeypatch the pipeline functions at their module boundary
+because rerun_embeddings is a script-level orchestrator and its
+testable contract is "calls the right functions with the right args".
+"""
+
+import scripts.rerun_embeddings as rerun
+import pipeline.text_pipeline as tp
+
+
+def test_rerun_retries_missing(monkeypatch):
+    """When ensure_text_embeddings returns failed_ids, retry helper is called."""
+    monkeypatch.setattr(rerun, "_clear_embeddings", lambda db_path: 0)
+    monkeypatch.setattr(rerun, "_get_all_windows", lambda db_path: [])
+
+    def first_call(db_path=None, model=None, batch_size=50, **kwargs):
+        return (1, 0, 0, 1, [101, 102])
+
+    called = {"retried": False, "ids": None}
+
+    def retry_call(db_path=None, ids=None, model=None, batch_size=10, **kwargs):
+        called["retried"] = True
+        called["ids"] = ids
+        return (1, 0, 0, 0, [])
+
+    monkeypatch.setattr(tp, "ensure_text_embeddings", first_call)
+    monkeypatch.setattr(tp, "ensure_text_embeddings_for_ids", retry_call)
+
+    summary = rerun.rerun_embeddings(
+        "data/motions.db", model="test-model", retry_missing=True
+    )
+
+    assert called["retried"] is True
+    assert set(called["ids"]) == {101, 102}
+
+
+def test_rerun_no_retry_when_no_failures(monkeypatch):
+    """When ensure_text_embeddings returns no failed_ids, retry is NOT called."""
+    monkeypatch.setattr(rerun, "_clear_embeddings", lambda db_path: 0)
+    monkeypatch.setattr(rerun, "_get_all_windows", lambda db_path: [])
+
+    def no_failures(db_path=None, model=None, batch_size=50, **kwargs):
+        return (5, 0, 0, 0, [])
+
+    retry_called = {"v": False}
+
+    def retry_should_not_be_called(**kwargs):
+        retry_called["v"] = True
+        return (0, 0, 0, 0, [])
+
+    monkeypatch.setattr(tp, "ensure_text_embeddings", no_failures)
+    monkeypatch.setattr(
+        tp, "ensure_text_embeddings_for_ids", retry_should_not_be_called
+    )
+
+    rerun.rerun_embeddings("data/motions.db", model="test-model", retry_missing=True)
+
+    assert retry_called["v"] is False
diff --git a/tests/test_similarity_compute_filter.py b/tests/test_similarity_compute_filter.py
new file mode 100644
index 0000000..1a8178c
--- /dev/null
+++ b/tests/test_similarity_compute_filter.py
@@ -0,0 +1,68 @@
+"""Tests for similarity filter in compute_similarities — real DB, real code, no mocks."""
+
+import json
+import duckdb
+from database import MotionDatabase
+import similarity.compute as sc
+
+
+def test_filter_skips_identical_short_title_pairs(tmp_path):
+    """Pairs with identical short titles and perfect cosine similarity are filtered out."""
+    db_path = str(tmp_path / "test.db")
+
+    # 1. Initialize schema
+    db = MotionDatabase(db_path)
+
+    # 2. Insert 2 motions with identical short titles
+    motion1 = {
+        "title": "Aangenomen.",
+        "description": "desc1",
+        "date": "2020-01-01",
+        "policy_area": "",
+        "voting_results": {},
+        "winning_margin": 0.5,
+        "url": "u1",
+    }
+    motion2 = {
+        "title": "Aangenomen.",
+        "description": "desc2",
+        "date": "2020-01-02",
+        "policy_area": "",
+        "voting_results": {},
+        "winning_margin": 0.6,
+        "url": "u2",
+    }
+
+    assert db.insert_motion(motion1) is True
+    assert db.insert_motion(motion2) is True
+
+    # fetch ids
+    conn = duckdb.connect(db_path)
+    id1 = conn.execute(
+        "SELECT id FROM motions WHERE url = ?", (motion1["url"],)
+    ).fetchone()[0]
+    id2 = conn.execute(
+        "SELECT id FROM motions WHERE url = ?", (motion2["url"],)
+    ).fetchone()[0]
+
+    assert id1 is not None and id2 is not None and id1 != id2
+
+    # 3. Insert identical unit vectors into fused_embeddings using store_fused_embedding
+    vec = [1.0] + [0.0] * 7  # 8-dim unit vector
+
+    # use a window id (schema requires NOT NULL); compute_similarities will read all fused embeddings when window_id=None
+    window_id = "w"
+    assert db.store_fused_embedding(id1, window_id, vec, svd_dims=0, text_dims=0) > 0
+    assert db.store_fused_embedding(id2, window_id, vec, svd_dims=0, text_dims=0) > 0
+    conn.close()
+
+    # 4. Run compute_similarities
+    inserted = sc.compute_similarities(
+        vector_type="fused",
+        window_id=None,
+        db_path=db_path,
+    )
+
+    # 5. The pair (id1, id2) has perfect similarity and identical short titles
+    #    The filter should remove it → 0 rows inserted into similarity_cache
+    assert inserted == 0, f"Expected 0 pairs after filter, got {inserted}"