@webship.co/workspace 1.0.0-beta1 → 1.0.0-beta2

package/core/scripts/libs/rag_index.py

@@ -0,0 +1,446 @@
+#!/usr/bin/env python3
+"""Index a project into Milvus, and read the result back.
+
+The worker behind cmd-tools-ragify.sh and cmd-tools-rag.sh. It is run through
+`uv run --with pymilvus ...` so there is nothing to install and nothing to keep
+in sync — see rag_python() in core/scripts/functions/fun-rag.sh.
+
+Retrieval is Milvus's built-in BM25 function by default: the database itself
+tokenises the text and builds the sparse vectors, so no embedding model runs
+here and nothing leaves the machine. `--embedding openai` adds a dense vector
+per chunk from the OpenAI API and searches both fields together, which is
+better on paraphrase and is the only mode that sends content anywhere.
+
+What gets indexed is deliberately the opposite of what graphify covers: prose,
+templates and configuration — .md, .twig, *.yml, *.module hooks — the files a
+code graph classifies as documents and then never reads.
+"""
+
+import argparse
+import json
+import os
+import sys
+from typing import Iterable, Iterator, List, Optional
+
+# Milvus VARCHAR tops out at 65535 bytes, and a chunk that long is not a useful
+# retrieval unit anyway. Chunks are truncated to this many characters, counted
+# in bytes after encoding because the limit is a byte limit and Arabic or emoji
+# in a comment would otherwise overflow a chunk that looks short.
+MAX_CHUNK_BYTES = 16000
+
+# Only these are read. A whitelist rather than a "skip binaries" heuristic: the
+# point is to index what a person would read, and a .png or a .lock is not that.
+TEXT_SUFFIXES = {
+    # Drupal/PHP, including the extensions that hold hooks — the files graphify
+    # does not classify at all.
+    ".php", ".module", ".install", ".theme", ".profile", ".engine", ".inc",
+    # The wiring and config a Drupal site actually behaves like.
+    ".yml", ".yaml", ".toml", ".ini", ".xml", ".json",
+    # Templates and front end.
+    ".twig", ".html", ".css", ".scss", ".less", ".js", ".jsx", ".ts", ".tsx", ".vue",
+    # Prose and scripts.
+    ".md", ".markdown", ".txt", ".rst", ".sh", ".bash", ".py", ".rb", ".go", ".sql",
+    ".feature", ".gitlab-ci.yml", ".dist",
+}
+
+# Files whose name alone says "read me" even without a suffix.
+TEXT_NAMES = {
+    "README", "CHANGELOG", "LICENSE", "AGENTS.md", "CLAUDE.md", "Dockerfile", "Makefile",
+    ".gitlab-ci.yml", "composer.json", "package.json",
+}
+
+# Never worth indexing even when the suffix passes: generated, minified or lock data.
+SKIP_NAME_PARTS = (".min.js", ".min.css", "-min.js", ".map", "composer.lock", "package-lock.json",
+                   "yarn.lock", "patches.lock.json")
+
+
+def eprint(*args):
+    print(*args, file=sys.stderr)
+
+
+def collection_ok(name: str) -> str:
+    """Milvus allows letters, digits and underscore, and no leading digit."""
+    safe = "".join(c if (c.isalnum() or c == "_") else "_" for c in name)
+    if safe and safe[0].isdigit():
+        safe = "c_" + safe
+    return safe
+
+
+def is_text_file(path: str) -> bool:
+    base = os.path.basename(path)
+    if any(part in base for part in SKIP_NAME_PARTS):
+        return False
+    if base in TEXT_NAMES:
+        return True
+    _, ext = os.path.splitext(base)
+    return ext.lower() in TEXT_SUFFIXES
+
+
+def walk_files(root: str, excludes: Iterable[str], max_file_kb: int) -> Iterator[str]:
+    """Every indexable file under root, minus the excluded paths.
+
+    Excludes are matched as path prefixes relative to the root ("web/core") and
+    as bare directory names ("node_modules"), which is how the workspace already
+    expresses them for graphify.
+    """
+    prefixes = {e.strip("/") for e in excludes if e and "/" in e}
+    names = {e.strip("/") for e in excludes if e and "/" not in e}
+    names |= {".git", ".ddev", ".idea", "graphify-out", "graphify-yaml"}
+    limit = max_file_kb * 1024
+
+    for dirpath, dirnames, filenames in os.walk(root):
+        rel_dir = os.path.relpath(dirpath, root)
+        rel_dir = "" if rel_dir == "." else rel_dir
+        # Prune in place so os.walk does not descend into what we are skipping.
+        dirnames[:] = [
+            d for d in dirnames
+            if d not in names
+            and os.path.join(rel_dir, d).lstrip("/") not in prefixes
+            and not any(os.path.join(rel_dir, d).lstrip("/").startswith(p + "/") for p in prefixes)
+        ]
+        for name in filenames:
+            full = os.path.join(dirpath, name)
+            if not is_text_file(full):
+                continue
+            try:
+                if os.path.getsize(full) > limit or os.path.islink(full):
+                    continue
+            except OSError:
+                continue
+            yield full
+
+
+def chunk_file(path: str, root: str, chunk_lines: int, overlap: int) -> List[dict]:
+    """Split one file into overlapping line windows.
+
+    Lines, not tokens: a chunk that keeps whole lines is one you can open at the
+    line number it reports, and the overlap keeps a function that straddles a
+    boundary findable from either side.
+    """
+    try:
+        with open(path, "r", encoding="utf-8", errors="replace") as handle:
+            lines = handle.read().splitlines()
+    except OSError:
+        return []
+
+    if not any(line.strip() for line in lines):
+        return []
+
+    rel = os.path.relpath(path, root)
+    ext = os.path.splitext(path)[1].lower().lstrip(".") or os.path.basename(path)
+    step = max(1, chunk_lines - overlap)
+    out = []
+
+    for start in range(0, max(1, len(lines)), step):
+        window = lines[start:start + chunk_lines]
+        text = "\n".join(window).strip()
+        if not text:
+            continue
+        encoded = text.encode("utf-8")
+        if len(encoded) > MAX_CHUNK_BYTES:
+            text = encoded[:MAX_CHUNK_BYTES].decode("utf-8", errors="ignore")
+        out.append({
+            "path": rel[:1020],
+            "ext": ext[:28],
+            "start_line": start + 1,
+            "end_line": min(start + chunk_lines, len(lines)),
+            # The path is part of the indexed text on purpose: "where is the
+            # Website Starter hero styled" should match a file called hero.twig even when
+            # the word never appears inside it.
+            "text": rel + "\n" + text,
+        })
+        if start + chunk_lines >= len(lines):
+            break
+    return out
+
+
+def connect(uri: str, token: str):
+    from pymilvus import MilvusClient
+    return MilvusClient(uri=uri, token=token or "")
+
+
+def openai_embed(texts: List[str], model: str, key: str) -> List[List[float]]:
+    from openai import OpenAI
+    client = OpenAI(api_key=key)
+    result = client.embeddings.create(model=model, input=texts)
+    return [item.embedding for item in result.data]
+
+
+def build_schema(client, embedding: str, dim: int):
+    from pymilvus import DataType, Function, FunctionType
+
+    schema = client.create_schema(auto_id=True, enable_dynamic_field=False)
+    schema.add_field("id", DataType.INT64, is_primary=True)
+    schema.add_field("path", DataType.VARCHAR, max_length=1024)
+    schema.add_field("ext", DataType.VARCHAR, max_length=32)
+    schema.add_field("start_line", DataType.INT64)
+    schema.add_field("end_line", DataType.INT64)
+    # enable_analyzer is what makes the BM25 function legal on this field: it is
+    # the tokeniser the sparse vectors are built from.
+    schema.add_field("text", DataType.VARCHAR, max_length=65535, enable_analyzer=True)
+    schema.add_field("sparse", DataType.SPARSE_FLOAT_VECTOR)
+    schema.add_function(Function(
+        name="text_bm25",
+        function_type=FunctionType.BM25,
+        input_field_names=["text"],
+        output_field_names=["sparse"],
+    ))
+
+    index_params = client.prepare_index_params()
+    index_params.add_index(
+        field_name="sparse",
+        index_type="SPARSE_INVERTED_INDEX",
+        metric_type="BM25",
+        params={"inverted_index_algo": "DAAT_MAXSCORE"},
+    )
+
+    if embedding == "openai":
+        schema.add_field("dense", DataType.FLOAT_VECTOR, dim=dim)
+        index_params.add_index(field_name="dense", index_type="AUTOINDEX", metric_type="COSINE")
+
+    return schema, index_params
+
+
+def cmd_index(args) -> int:
+    root = os.path.abspath(args.path)
+    if not os.path.isdir(root):
+        eprint(f"No such project directory: {root}")
+        return 1
+
+    name = collection_ok(args.collection)
+    embedding = args.embedding
+    key = args.openai_key or os.environ.get("OPENAI_API_KEY", "")
+    if embedding == "openai" and not key:
+        eprint("--embedding openai needs an OpenAI key (settings.yml ai_providers.openai.api_key "
+               "or OPENAI_API_KEY).")
+        return 1
+
+    client = connect(args.uri, args.token)
+
+    # A dimension is only known once a model has answered, so it is probed with one
+    # cheap call rather than hardcoded per model name.
+    dim = 0
+    if embedding == "openai":
+        dim = len(openai_embed(["dimension probe"], args.openai_model, key)[0])
+
+    existing = name in client.list_collections()
+    if existing and not args.append:
+        client.drop_collection(name)
+        existing = False
+    if not existing:
+        schema, index_params = build_schema(client, embedding, dim)
+        client.create_collection(collection_name=name, schema=schema, index_params=index_params)
+
+    files = list(walk_files(root, args.exclude or [], args.max_file_kb))
+    total_chunks = 0
+    batch: List[dict] = []
+    batch_size = 128 if embedding == "openai" else 512
+
+    def flush() -> int:
+        nonlocal batch
+        if not batch:
+            return 0
+        if embedding == "openai":
+            vectors = openai_embed([row["text"] for row in batch], args.openai_model, key)
+            for row, vector in zip(batch, vectors):
+                row["dense"] = vector
+        client.insert(collection_name=name, data=batch)
+        count = len(batch)
+        batch = []
+        return count
+
+    for index, path in enumerate(files, start=1):
+        for chunk in chunk_file(path, root, args.chunk_lines, args.chunk_overlap):
+            batch.append(chunk)
+            if len(batch) >= batch_size:
+                total_chunks += flush()
+        if args.progress and index % 200 == 0:
+            eprint(f"  {index}/{len(files)} files, {total_chunks + len(batch)} chunks")
+
+    total_chunks += flush()
+    client.flush(name)
+    client.load_collection(name)
+
+    result = {
+        "collection": name,
+        "files": len(files),
+        "chunks": total_chunks,
+        "embedding": embedding,
+        "uri": args.uri,
+    }
+    print(json.dumps(result) if args.json else
+          f"Indexed {result['files']} files as {result['chunks']} chunks "
+          f"into {name} ({embedding}).")
+    return 0
+
+
+def cmd_search(args) -> int:
+    client = connect(args.uri, args.token)
+    name = collection_ok(args.collection)
+    if name not in client.list_collections():
+        eprint(f"No collection {name}. Index it first with cmd-tools-ragify.sh.")
+        return 1
+    client.load_collection(name)
+
+    fields = ["path", "start_line", "end_line", "text"]
+    has_dense = any(f["name"] == "dense" for f in client.describe_collection(name)["fields"])
+
+    if has_dense and args.openai_key:
+        # Dense and sparse are asked separately and merged by reciprocal rank
+        # fusion: BM25 finds the exact identifier, the embedding finds the
+        # paraphrase, and RRF needs no score calibration between the two.
+        from pymilvus import AnnSearchRequest, RRFRanker
+        vector = openai_embed([args.query], args.openai_model, args.openai_key)[0]
+        hits = client.hybrid_search(
+            collection_name=name,
+            reqs=[
+                AnnSearchRequest(data=[args.query], anns_field="sparse", param={}, limit=args.limit),
+                AnnSearchRequest(data=[vector], anns_field="dense", param={}, limit=args.limit),
+            ],
+            ranker=RRFRanker(),
+            limit=args.limit,
+            output_fields=fields,
+        )
+    else:
+        hits = client.search(
+            collection_name=name,
+            data=[args.query],
+            anns_field="sparse",
+            limit=args.limit,
+            output_fields=fields,
+        )
+
+    rows = [
+        {
+            "score": round(float(hit["distance"]), 4),
+            "path": hit["entity"]["path"],
+            "lines": f"{hit['entity']['start_line']}-{hit['entity']['end_line']}",
+            "text": hit["entity"]["text"],
+        }
+        for hit in (hits[0] if hits else [])
+    ]
+
+    if args.json:
+        print(json.dumps(rows, indent=2))
+        return 0
+
+    if not rows:
+        print(f"No match for {args.query!r} in {name}.")
+        return 0
+
+    for row in rows:
+        print(f"\n\033[1m{row['path']}:{row['lines']}\033[0m  (score {row['score']})")
+        body = row["text"].split("\n", 1)[1] if "\n" in row["text"] else row["text"]
+        for line in body.splitlines()[: args.context]:
+            print(f"  {line}")
+    print()
+    return 0
+
+
+def cmd_collections(args) -> int:
+    client = connect(args.uri, args.token)
+    names = sorted(client.list_collections())
+    rows = []
+    for name in names:
+        try:
+            stats = client.get_collection_stats(name)
+            count = int(stats.get("row_count", 0))
+        except Exception:
+            count = -1
+        rows.append({"collection": name, "chunks": count})
+
+    if args.json:
+        print(json.dumps(rows, indent=2))
+        return 0
+    if not rows:
+        print("No collections yet. Index one with: cmd-tools-ragify.sh <project>")
+        return 0
+    print(f"{'COLLECTION':<52} {'CHUNKS':>10}")
+    for row in rows:
+        print(f"{row['collection']:<52} {row['chunks']:>10}")
+    return 0
+
+
+def cmd_info(args) -> int:
+    client = connect(args.uri, args.token)
+    name = collection_ok(args.collection)
+    if name not in client.list_collections():
+        eprint(f"No collection {name}.")
+        return 1
+    described = client.describe_collection(name)
+    stats = client.get_collection_stats(name)
+    info = {
+        "collection": name,
+        "chunks": int(stats.get("row_count", 0)),
+        "fields": [f["name"] for f in described["fields"]],
+        "functions": [f.get("name") for f in described.get("functions", [])],
+    }
+    if args.json:
+        print(json.dumps(info, indent=2))
+        return 0
+    print(f"Collection : {info['collection']}")
+    print(f"Chunks     : {info['chunks']}")
+    print(f"Fields     : {', '.join(info['fields'])}")
+    print(f"Functions  : {', '.join(x for x in info['functions'] if x) or '—'}")
+    return 0
+
+
+def cmd_drop(args) -> int:
+    client = connect(args.uri, args.token)
+    name = collection_ok(args.collection)
+    if name not in client.list_collections():
+        eprint(f"No collection {name}.")
+        return 1
+    client.drop_collection(name)
+    print(f"Dropped {name}. Rebuild it with cmd-tools-ragify.sh in its workspace folder.")
+    return 0
+
+
+def main(argv: Optional[List[str]] = None) -> int:
+    parser = argparse.ArgumentParser(description="Index a project into Milvus and query it.")
+    parser.add_argument("--uri", default="http://127.0.0.1:19530")
+    parser.add_argument("--token", default="")
+    parser.add_argument("--json", action="store_true")
+    sub = parser.add_subparsers(dest="command", required=True)
+
+    p = sub.add_parser("index")
+    p.add_argument("--collection", required=True)
+    p.add_argument("--path", required=True)
+    p.add_argument("--exclude", nargs="*", default=[])
+    p.add_argument("--chunk-lines", type=int, default=60)
+    p.add_argument("--chunk-overlap", type=int, default=10)
+    p.add_argument("--max-file-kb", type=int, default=512)
+    p.add_argument("--embedding", choices=["bm25", "openai"], default="bm25")
+    p.add_argument("--openai-model", default="text-embedding-3-small")
+    p.add_argument("--openai-key", default="")
+    p.add_argument("--append", action="store_true",
+                   help="Add to the collection instead of rebuilding it.")
+    p.add_argument("--progress", action="store_true")
+    p.set_defaults(func=cmd_index)
+
+    p = sub.add_parser("search")
+    p.add_argument("--collection", required=True)
+    p.add_argument("--query", required=True)
+    p.add_argument("--limit", type=int, default=5)
+    p.add_argument("--context", type=int, default=12)
+    p.add_argument("--openai-model", default="text-embedding-3-small")
+    p.add_argument("--openai-key", default="")
+    p.set_defaults(func=cmd_search)
+
+    p = sub.add_parser("collections")
+    p.set_defaults(func=cmd_collections)
+
+    p = sub.add_parser("info")
+    p.add_argument("--collection", required=True)
+    p.set_defaults(func=cmd_info)
+
+    p = sub.add_parser("drop")
+    p.add_argument("--collection", required=True)
+    p.set_defaults(func=cmd_drop)
+
+    args = parser.parse_args(argv)
+    return args.func(args)
+
+
+if __name__ == "__main__":
+    sys.exit(main())