feed-the-machine 1.6.1 → 1.7.1

package/ftm-map/scripts/ranker.py

@@ -1,377 +1,377 @@
-"""
-ranker.py -- PageRank-based context selection engine for ftm-map.
-
-Implements Aider-style personalized PageRank over the file-level dependency graph
-with task-aware personalization and token-budget binary search.
-"""
-
-import math
-import os
-import sys
-
-sys.path.insert(0, os.path.dirname(__file__))
-
-import numpy as np
-import scipy.sparse as sp
-
-# Try fast-pagerank first, fall back to scipy power iteration
-try:
-    from fast_pagerank import pagerank_power
-    HAS_FAST_PAGERANK = True
-except ImportError:
-    HAS_FAST_PAGERANK = False
-
-
-def build_adjacency_matrix(conn):
-    """Build undirected sparse adjacency matrix from file_edges.
-
-    Returns (matrix, file_id_to_idx, idx_to_file_id) where:
-    - matrix is a scipy CSR sparse matrix (undirected: A + A.T)
-    - file_id_to_idx maps file_id -> matrix index
-    - idx_to_file_id maps matrix index -> file_id
-    """
-    # Get all files
-    files = conn.execute("SELECT id FROM files ORDER BY id").fetchall()
-    if not files:
-        return None, {}, {}
-
-    file_ids = [row['id'] for row in files]
-    file_id_to_idx = {fid: i for i, fid in enumerate(file_ids)}
-    idx_to_file_id = {i: fid for i, fid in enumerate(file_ids)}
-    n = len(file_ids)
-
-    # Get edges
-    edges = conn.execute(
-        "SELECT source_file_id, target_file_id, weight FROM file_edges"
-    ).fetchall()
-
-    if not edges:
-        return sp.csr_matrix((n, n)), file_id_to_idx, idx_to_file_id
-
-    rows, cols, data = [], [], []
-    for edge in edges:
-        src_idx = file_id_to_idx.get(edge['source_file_id'])
-        tgt_idx = file_id_to_idx.get(edge['target_file_id'])
-        if src_idx is not None and tgt_idx is not None:
-            rows.append(src_idx)
-            cols.append(tgt_idx)
-            data.append(edge['weight'])
-
-    # Build directed matrix, then symmetrize for undirected PageRank
-    A = sp.csr_matrix((data, (rows, cols)), shape=(n, n))
-    A_undirected = A + A.T  # Symmetrize
-
-    return A_undirected, file_id_to_idx, idx_to_file_id
-
-
-def build_personalization(
-    conn, seed_files=None, seed_keywords=None, seed_symbols=None, file_id_to_idx=None
-):
-    """Build personalization vector for PageRank.
-
-    Three channels:
-    - seed_files: file paths get 100x weight
-    - seed_keywords: FTS5 matches get 30x weight
-    - seed_symbols: symbol name matches - defining file gets 80x, referencing files get 40x
-
-    Returns normalized numpy array (sums to 1.0).
-    """
-    n = len(file_id_to_idx)
-    if n == 0:
-        return None
-
-    pers = np.ones(n)  # Base: uniform weight of 1
-
-    # Channel 1: Seed files (100x)
-    if seed_files:
-        for fpath in seed_files:
-            file_row = conn.execute(
-                "SELECT id FROM files WHERE path=?", (fpath,)
-            ).fetchone()
-            if file_row and file_row['id'] in file_id_to_idx:
-                idx = file_id_to_idx[file_row['id']]
-                pers[idx] *= 100
-
-    # Channel 2: Seed keywords via FTS5 (30x)
-    if seed_keywords:
-        for kw in seed_keywords:
-            try:
-                fts_results = conn.execute(
-                    "SELECT s.file_id FROM symbols_fts fts "
-                    "JOIN symbols s ON s.id = fts.rowid "
-                    "WHERE symbols_fts MATCH ? LIMIT 50",
-                    (kw,),
-                ).fetchall()
-                for row in fts_results:
-                    if row['file_id'] in file_id_to_idx:
-                        pers[file_id_to_idx[row['file_id']]] *= 30
-            except Exception:
-                pass  # FTS query syntax errors are non-fatal
-
-    # Channel 3: Seed symbols (80x defining, 40x referencing)
-    if seed_symbols:
-        for sym_name in seed_symbols:
-            # Defining files get 80x
-            def_files = conn.execute(
-                "SELECT DISTINCT file_id FROM symbols WHERE name=?", (sym_name,)
-            ).fetchall()
-            for row in def_files:
-                if row['file_id'] in file_id_to_idx:
-                    pers[file_id_to_idx[row['file_id']]] *= 80
-
-            # Referencing files get 40x
-            ref_files = conn.execute(
-                "SELECT DISTINCT file_id FROM refs WHERE symbol_name=?", (sym_name,)
-            ).fetchall()
-            for row in ref_files:
-                if row['file_id'] in file_id_to_idx:
-                    pers[file_id_to_idx[row['file_id']]] *= 40
-
-    # Normalize to sum to 1
-    total = pers.sum()
-    if total > 0:
-        pers /= total
-
-    return pers
-
-
-def run_pagerank(adj_matrix, personalization=None, damping=0.85, max_iter=100, tol=1e-6):
-    """Run PageRank on the adjacency matrix.
-
-    Uses fast-pagerank if available, otherwise scipy power iteration.
-    Returns numpy array of scores indexed by matrix position.
-    """
-    n = adj_matrix.shape[0]
-    if n == 0:
-        return np.array([])
-
-    if HAS_FAST_PAGERANK and personalization is not None:
-        try:
-            scores = pagerank_power(
-                adj_matrix, p=damping, personalize=personalization, tol=tol
-            )
-            return scores
-        except Exception:
-            pass  # Fall through to scipy implementation
-
-    # Scipy power iteration fallback
-    # Normalize adjacency matrix columns (column-stochastic transition matrix)
-    col_sums = np.array(adj_matrix.sum(axis=0)).flatten()
-    col_sums[col_sums == 0] = 1  # Avoid division by zero for dangling nodes
-
-    # Transition matrix: M[i,j] = A[i,j] / col_sum[j]
-    D_inv = sp.diags(1.0 / col_sums)
-    M = adj_matrix @ D_inv
-
-    # Initialize personalization / teleport vector
-    if personalization is not None:
-        v = personalization.copy()
-    else:
-        v = np.ones(n) / n
-
-    scores = v.copy()
-
-    # Dangling nodes: columns with zero outgoing weight
-    dangling_mask = np.array(adj_matrix.sum(axis=0)).flatten() == 0
-
-    for _ in range(max_iter):
-        prev = scores.copy()
-
-        # PageRank iteration with dangling-node redistribution
-        dangling_sum = scores[dangling_mask].sum() if dangling_mask.any() else 0
-        scores = damping * (M @ scores) + damping * dangling_sum * v + (1 - damping) * v
-
-        # Check convergence via L1 norm
-        if np.abs(scores - prev).sum() < tol:
-            break
-
-    return scores
-
-
-def rank_files(conn, seed_files=None, seed_keywords=None, seed_symbols=None):
-    """Rank all files by structural importance with personalization.
-
-    Returns sorted list of (file_path, score) tuples, highest score first.
-    """
-    adj, fid_to_idx, idx_to_fid = build_adjacency_matrix(conn)
-    if adj is None or adj.shape[0] == 0:
-        return []
-
-    pers = build_personalization(
-        conn, seed_files, seed_keywords, seed_symbols, fid_to_idx
-    )
-    scores = run_pagerank(adj, pers)
-
-    # Map scores back to file paths
-    results = []
-    for idx, score in enumerate(scores):
-        file_id = idx_to_fid[idx]
-        file_row = conn.execute(
-            "SELECT path FROM files WHERE id=?", (file_id,)
-        ).fetchone()
-        if file_row:
-            results.append((file_row['path'], float(score)))
-
-    results.sort(key=lambda x: x[1], reverse=True)
-    return results
-
-
-def fit_to_budget(ranked_files, conn, token_budget):
-    """Select files + key symbols that fit within token budget.
-
-    Uses binary search with 15% tolerance (Aider's approach).
-    Token estimation: ~25 tokens per tag/symbol entry.
-
-    Returns (result_list, total_tokens) where result_list contains dicts:
-        [{path, score, symbols: [name, ...], tokens}]
-    """
-    if not ranked_files or token_budget <= 0:
-        return [], 0
-
-    def estimate_tokens(file_list):
-        """Estimate tokens for a list of files based on their symbol count."""
-        total = 0
-        for fpath, _ in file_list:
-            file_row = conn.execute(
-                "SELECT id, line_count FROM files WHERE path=?", (fpath,)
-            ).fetchone()
-            if not file_row:
-                continue
-            syms = conn.execute(
-                "SELECT name, signature FROM symbols WHERE file_id=? ORDER BY line_start",
-                (file_row['id'],),
-            ).fetchall()
-            for _sym in syms:
-                # ~25 tokens per tag entry (Aider's estimate)
-                total += 25
-        return total
-
-    # Binary search: find max number of files that fits within budget
-    lo, hi = 1, len(ranked_files)
-    best = 1
-
-    while lo <= hi:
-        mid = (lo + hi) // 2
-        tokens = estimate_tokens(ranked_files[:mid])
-        if tokens <= token_budget:
-            best = mid
-            lo = mid + 1
-        else:
-            hi = mid - 1
-
-    # Allow 15% tolerance -- greedily add more files if within tolerance
-    tolerance_budget = token_budget * 1.15
-    while best < len(ranked_files):
-        tokens = estimate_tokens(ranked_files[: best + 1])
-        if tokens <= tolerance_budget:
-            best += 1
-        else:
-            break
-
-    # Build output with symbols for each selected file
-    result = []
-    total_tokens = 0
-    for fpath, score in ranked_files[:best]:
-        file_row = conn.execute(
-            "SELECT id FROM files WHERE path=?", (fpath,)
-        ).fetchone()
-        if not file_row:
-            continue
-        syms = conn.execute(
-            "SELECT name FROM symbols WHERE file_id=? ORDER BY line_start",
-            (file_row['id'],),
-        ).fetchall()
-        sym_names = [s['name'] for s in syms]
-        entry_tokens = len(sym_names) * 25
-        total_tokens += entry_tokens
-        result.append({
-            "path": fpath,
-            "score": round(score, 6),
-            "symbols": sym_names,
-            "tokens": entry_tokens,
-        })
-
-    return result, total_tokens
-
-
-# ---------------------------------------------------------------------------
-# Smoke test
-# ---------------------------------------------------------------------------
-
-if __name__ == "__main__":
-    import tempfile
-    from db import (
-        get_connection,
-        add_file,
-        add_symbol,
-        add_reference,
-        rebuild_file_edges,
-        rebuild_symbol_edges,
-    )
-
-    print("Running ranker.py smoke tests ...")
-
-    with tempfile.TemporaryDirectory() as tmp:
-        conn = get_connection(tmp)
-
-        # Create a small graph: 3 files with cross-references
-        f1 = add_file(conn, "src/auth.py", "python", 1.0, line_count=50)
-        f2 = add_file(conn, "src/api.py", "python", 1.0, line_count=100)
-        f3 = add_file(conn, "src/utils.py", "python", 1.0, line_count=30)
-
-        # Symbols
-        add_symbol(
-            conn, f1, "authenticate", "definition", 1, 20,
-            signature="def authenticate(req)",
-        )
-        add_symbol(conn, f1, "verify_token", "definition", 25, 40)
-        add_symbol(
-            conn, f2, "handle_request", "definition", 1, 50,
-            signature="def handle_request(req)",
-        )
-        add_symbol(conn, f3, "format_date", "definition", 1, 10)
-        add_symbol(conn, f3, "parse_config", "definition", 15, 25)
-
-        # References: api.py references auth.py functions, and utils.py
-        add_reference(conn, f2, "authenticate", 10)
-        add_reference(conn, f2, "verify_token", 15)
-        add_reference(conn, f2, "format_date", 20)
-        add_reference(conn, f2, "parse_config", 25)
-        # auth.py also references utils
-        add_reference(conn, f1, "parse_config", 30)
-
-        # Materialize edges
-        rebuild_file_edges(conn)
-        conn.commit()
-
-        # Test 1: Uniform PageRank
-        results = rank_files(conn)
-        print(f"  Uniform PageRank: {len(results)} files ranked")
-        for path, score in results:
-            print(f"    {path}: {score:.6f}")
-        assert len(results) == 3
-
-        # Test 2: Personalized -- seed auth.py
-        results_pers = rank_files(conn, seed_files=["src/auth.py"])
-        print(f"  Personalized (seed auth.py): {len(results_pers)} files")
-        for path, score in results_pers:
-            print(f"    {path}: {score:.6f}")
-        # auth.py should be ranked higher with personalization
-        auth_score = next(s for p, s in results_pers if p == "src/auth.py")
-        auth_uniform = next(s for p, s in results if p == "src/auth.py")
-        print(f"  Auth personalized boost: {auth_score:.6f} vs {auth_uniform:.6f}")
-
-        # Test 3: Budget fitting
-        budget_result, total_tokens = fit_to_budget(results, conn, 200)
-        print(f"  Budget fit (200 tokens): {len(budget_result)} files, {total_tokens} tokens")
-        assert total_tokens <= 200 * 1.15  # 15% tolerance
-
-        # Test 4: Keyword personalization
-        results_kw = rank_files(conn, seed_keywords=["authenticate"])
-        print(f"  Keyword personalized: {len(results_kw)} files")
-
-        # Test 5: Symbol personalization
-        results_sym = rank_files(conn, seed_symbols=["authenticate"])
-        print(f"  Symbol personalized: {len(results_sym)} files")
-
-        print("\nAll ranker smoke tests passed.")
+"""
+ranker.py -- PageRank-based context selection engine for ftm-map.
+
+Implements Aider-style personalized PageRank over the file-level dependency graph
+with task-aware personalization and token-budget binary search.
+"""
+
+import math
+import os
+import sys
+
+sys.path.insert(0, os.path.dirname(__file__))
+
+import numpy as np
+import scipy.sparse as sp
+
+# Try fast-pagerank first, fall back to scipy power iteration
+try:
+    from fast_pagerank import pagerank_power
+    HAS_FAST_PAGERANK = True
+except ImportError:
+    HAS_FAST_PAGERANK = False
+
+
+def build_adjacency_matrix(conn):
+    """Build undirected sparse adjacency matrix from file_edges.
+
+    Returns (matrix, file_id_to_idx, idx_to_file_id) where:
+    - matrix is a scipy CSR sparse matrix (undirected: A + A.T)
+    - file_id_to_idx maps file_id -> matrix index
+    - idx_to_file_id maps matrix index -> file_id
+    """
+    # Get all files
+    files = conn.execute("SELECT id FROM files ORDER BY id").fetchall()
+    if not files:
+        return None, {}, {}
+
+    file_ids = [row['id'] for row in files]
+    file_id_to_idx = {fid: i for i, fid in enumerate(file_ids)}
+    idx_to_file_id = {i: fid for i, fid in enumerate(file_ids)}
+    n = len(file_ids)
+
+    # Get edges
+    edges = conn.execute(
+        "SELECT source_file_id, target_file_id, weight FROM file_edges"
+    ).fetchall()
+
+    if not edges:
+        return sp.csr_matrix((n, n)), file_id_to_idx, idx_to_file_id
+
+    rows, cols, data = [], [], []
+    for edge in edges:
+        src_idx = file_id_to_idx.get(edge['source_file_id'])
+        tgt_idx = file_id_to_idx.get(edge['target_file_id'])
+        if src_idx is not None and tgt_idx is not None:
+            rows.append(src_idx)
+            cols.append(tgt_idx)
+            data.append(edge['weight'])
+
+    # Build directed matrix, then symmetrize for undirected PageRank
+    A = sp.csr_matrix((data, (rows, cols)), shape=(n, n))
+    A_undirected = A + A.T  # Symmetrize
+
+    return A_undirected, file_id_to_idx, idx_to_file_id
+
+
+def build_personalization(
+    conn, seed_files=None, seed_keywords=None, seed_symbols=None, file_id_to_idx=None
+):
+    """Build personalization vector for PageRank.
+
+    Three channels:
+    - seed_files: file paths get 100x weight
+    - seed_keywords: FTS5 matches get 30x weight
+    - seed_symbols: symbol name matches - defining file gets 80x, referencing files get 40x
+
+    Returns normalized numpy array (sums to 1.0).
+    """
+    n = len(file_id_to_idx)
+    if n == 0:
+        return None
+
+    pers = np.ones(n)  # Base: uniform weight of 1
+
+    # Channel 1: Seed files (100x)
+    if seed_files:
+        for fpath in seed_files:
+            file_row = conn.execute(
+                "SELECT id FROM files WHERE path=?", (fpath,)
+            ).fetchone()
+            if file_row and file_row['id'] in file_id_to_idx:
+                idx = file_id_to_idx[file_row['id']]
+                pers[idx] *= 100
+
+    # Channel 2: Seed keywords via FTS5 (30x)
+    if seed_keywords:
+        for kw in seed_keywords:
+            try:
+                fts_results = conn.execute(
+                    "SELECT s.file_id FROM symbols_fts fts "
+                    "JOIN symbols s ON s.id = fts.rowid "
+                    "WHERE symbols_fts MATCH ? LIMIT 50",
+                    (kw,),
+                ).fetchall()
+                for row in fts_results:
+                    if row['file_id'] in file_id_to_idx:
+                        pers[file_id_to_idx[row['file_id']]] *= 30
+            except Exception:
+                pass  # FTS query syntax errors are non-fatal
+
+    # Channel 3: Seed symbols (80x defining, 40x referencing)
+    if seed_symbols:
+        for sym_name in seed_symbols:
+            # Defining files get 80x
+            def_files = conn.execute(
+                "SELECT DISTINCT file_id FROM symbols WHERE name=?", (sym_name,)
+            ).fetchall()
+            for row in def_files:
+                if row['file_id'] in file_id_to_idx:
+                    pers[file_id_to_idx[row['file_id']]] *= 80
+
+            # Referencing files get 40x
+            ref_files = conn.execute(
+                "SELECT DISTINCT file_id FROM refs WHERE symbol_name=?", (sym_name,)
+            ).fetchall()
+            for row in ref_files:
+                if row['file_id'] in file_id_to_idx:
+                    pers[file_id_to_idx[row['file_id']]] *= 40
+
+    # Normalize to sum to 1
+    total = pers.sum()
+    if total > 0:
+        pers /= total
+
+    return pers
+
+
+def run_pagerank(adj_matrix, personalization=None, damping=0.85, max_iter=100, tol=1e-6):
+    """Run PageRank on the adjacency matrix.
+
+    Uses fast-pagerank if available, otherwise scipy power iteration.
+    Returns numpy array of scores indexed by matrix position.
+    """
+    n = adj_matrix.shape[0]
+    if n == 0:
+        return np.array([])
+
+    if HAS_FAST_PAGERANK and personalization is not None:
+        try:
+            scores = pagerank_power(
+                adj_matrix, p=damping, personalize=personalization, tol=tol
+            )
+            return scores
+        except Exception:
+            pass  # Fall through to scipy implementation
+
+    # Scipy power iteration fallback
+    # Normalize adjacency matrix columns (column-stochastic transition matrix)
+    col_sums = np.array(adj_matrix.sum(axis=0)).flatten()
+    col_sums[col_sums == 0] = 1  # Avoid division by zero for dangling nodes
+
+    # Transition matrix: M[i,j] = A[i,j] / col_sum[j]
+    D_inv = sp.diags(1.0 / col_sums)
+    M = adj_matrix @ D_inv
+
+    # Initialize personalization / teleport vector
+    if personalization is not None:
+        v = personalization.copy()
+    else:
+        v = np.ones(n) / n
+
+    scores = v.copy()
+
+    # Dangling nodes: columns with zero outgoing weight
+    dangling_mask = np.array(adj_matrix.sum(axis=0)).flatten() == 0
+
+    for _ in range(max_iter):
+        prev = scores.copy()
+
+        # PageRank iteration with dangling-node redistribution
+        dangling_sum = scores[dangling_mask].sum() if dangling_mask.any() else 0
+        scores = damping * (M @ scores) + damping * dangling_sum * v + (1 - damping) * v
+
+        # Check convergence via L1 norm
+        if np.abs(scores - prev).sum() < tol:
+            break
+
+    return scores
+
+
+def rank_files(conn, seed_files=None, seed_keywords=None, seed_symbols=None):
+    """Rank all files by structural importance with personalization.
+
+    Returns sorted list of (file_path, score) tuples, highest score first.
+    """
+    adj, fid_to_idx, idx_to_fid = build_adjacency_matrix(conn)
+    if adj is None or adj.shape[0] == 0:
+        return []
+
+    pers = build_personalization(
+        conn, seed_files, seed_keywords, seed_symbols, fid_to_idx
+    )
+    scores = run_pagerank(adj, pers)
+
+    # Map scores back to file paths
+    results = []
+    for idx, score in enumerate(scores):
+        file_id = idx_to_fid[idx]
+        file_row = conn.execute(
+            "SELECT path FROM files WHERE id=?", (file_id,)
+        ).fetchone()
+        if file_row:
+            results.append((file_row['path'], float(score)))
+
+    results.sort(key=lambda x: x[1], reverse=True)
+    return results
+
+
+def fit_to_budget(ranked_files, conn, token_budget):
+    """Select files + key symbols that fit within token budget.
+
+    Uses binary search with 15% tolerance (Aider's approach).
+    Token estimation: ~25 tokens per tag/symbol entry.
+
+    Returns (result_list, total_tokens) where result_list contains dicts:
+        [{path, score, symbols: [name, ...], tokens}]
+    """
+    if not ranked_files or token_budget <= 0:
+        return [], 0
+
+    def estimate_tokens(file_list):
+        """Estimate tokens for a list of files based on their symbol count."""
+        total = 0
+        for fpath, _ in file_list:
+            file_row = conn.execute(
+                "SELECT id, line_count FROM files WHERE path=?", (fpath,)
+            ).fetchone()
+            if not file_row:
+                continue
+            syms = conn.execute(
+                "SELECT name, signature FROM symbols WHERE file_id=? ORDER BY line_start",
+                (file_row['id'],),
+            ).fetchall()
+            for _sym in syms:
+                # ~25 tokens per tag entry (Aider's estimate)
+                total += 25
+        return total
+
+    # Binary search: find max number of files that fits within budget
+    lo, hi = 1, len(ranked_files)
+    best = 1
+
+    while lo <= hi:
+        mid = (lo + hi) // 2
+        tokens = estimate_tokens(ranked_files[:mid])
+        if tokens <= token_budget:
+            best = mid
+            lo = mid + 1
+        else:
+            hi = mid - 1
+
+    # Allow 15% tolerance -- greedily add more files if within tolerance
+    tolerance_budget = token_budget * 1.15
+    while best < len(ranked_files):
+        tokens = estimate_tokens(ranked_files[: best + 1])
+        if tokens <= tolerance_budget:
+            best += 1
+        else:
+            break
+
+    # Build output with symbols for each selected file
+    result = []
+    total_tokens = 0
+    for fpath, score in ranked_files[:best]:
+        file_row = conn.execute(
+            "SELECT id FROM files WHERE path=?", (fpath,)
+        ).fetchone()
+        if not file_row:
+            continue
+        syms = conn.execute(
+            "SELECT name FROM symbols WHERE file_id=? ORDER BY line_start",
+            (file_row['id'],),
+        ).fetchall()
+        sym_names = [s['name'] for s in syms]
+        entry_tokens = len(sym_names) * 25
+        total_tokens += entry_tokens
+        result.append({
+            "path": fpath,
+            "score": round(score, 6),
+            "symbols": sym_names,
+            "tokens": entry_tokens,
+        })
+
+    return result, total_tokens
+
+
+# ---------------------------------------------------------------------------
+# Smoke test
+# ---------------------------------------------------------------------------
+
+if __name__ == "__main__":
+    import tempfile
+    from db import (
+        get_connection,
+        add_file,
+        add_symbol,
+        add_reference,
+        rebuild_file_edges,
+        rebuild_symbol_edges,
+    )
+
+    print("Running ranker.py smoke tests ...")
+
+    with tempfile.TemporaryDirectory() as tmp:
+        conn = get_connection(tmp)
+
+        # Create a small graph: 3 files with cross-references
+        f1 = add_file(conn, "src/auth.py", "python", 1.0, line_count=50)
+        f2 = add_file(conn, "src/api.py", "python", 1.0, line_count=100)
+        f3 = add_file(conn, "src/utils.py", "python", 1.0, line_count=30)
+
+        # Symbols
+        add_symbol(
+            conn, f1, "authenticate", "definition", 1, 20,
+            signature="def authenticate(req)",
+        )
+        add_symbol(conn, f1, "verify_token", "definition", 25, 40)
+        add_symbol(
+            conn, f2, "handle_request", "definition", 1, 50,
+            signature="def handle_request(req)",
+        )
+        add_symbol(conn, f3, "format_date", "definition", 1, 10)
+        add_symbol(conn, f3, "parse_config", "definition", 15, 25)
+
+        # References: api.py references auth.py functions, and utils.py
+        add_reference(conn, f2, "authenticate", 10)
+        add_reference(conn, f2, "verify_token", 15)
+        add_reference(conn, f2, "format_date", 20)
+        add_reference(conn, f2, "parse_config", 25)
+        # auth.py also references utils
+        add_reference(conn, f1, "parse_config", 30)
+
+        # Materialize edges
+        rebuild_file_edges(conn)
+        conn.commit()
+
+        # Test 1: Uniform PageRank
+        results = rank_files(conn)
+        print(f"  Uniform PageRank: {len(results)} files ranked")
+        for path, score in results:
+            print(f"    {path}: {score:.6f}")
+        assert len(results) == 3
+
+        # Test 2: Personalized -- seed auth.py
+        results_pers = rank_files(conn, seed_files=["src/auth.py"])
+        print(f"  Personalized (seed auth.py): {len(results_pers)} files")
+        for path, score in results_pers:
+            print(f"    {path}: {score:.6f}")
+        # auth.py should be ranked higher with personalization
+        auth_score = next(s for p, s in results_pers if p == "src/auth.py")
+        auth_uniform = next(s for p, s in results if p == "src/auth.py")
+        print(f"  Auth personalized boost: {auth_score:.6f} vs {auth_uniform:.6f}")
+
+        # Test 3: Budget fitting
+        budget_result, total_tokens = fit_to_budget(results, conn, 200)
+        print(f"  Budget fit (200 tokens): {len(budget_result)} files, {total_tokens} tokens")
+        assert total_tokens <= 200 * 1.15  # 15% tolerance
+
+        # Test 4: Keyword personalization
+        results_kw = rank_files(conn, seed_keywords=["authenticate"])
+        print(f"  Keyword personalized: {len(results_kw)} files")
+
+        # Test 5: Symbol personalization
+        results_sym = rank_files(conn, seed_symbols=["authenticate"])
+        print(f"  Symbol personalized: {len(results_sym)} files")
+
+        print("\nAll ranker smoke tests passed.")