Initial SQL-only 1C adapter baseline

2026-07-22 03:03:47 +03:00
commit e2503b77e7
545 changed files with 184711 additions and 0 deletions
@@ -0,0 +1,114 @@
+from __future__ import annotations
+
+import argparse
+import sys
+import urllib.error
+from pathlib import Path
+
+from common import call_chat_completion, read_json, search_lexical_index
+from rag_profiles import resolve_rag_profile
+
+ROOT = Path(__file__).resolve().parents[1]
+DEFAULT_INDEX = ROOT / "plugins" / "1c" / "datasets" / "prepared" / "rag_index.json"
+DEFAULT_SYSTEM_PROMPT = ROOT / "plugins" / "1c" / "prompts" / "system.md"
+DEFAULT_RAG_PROMPT = ROOT / "plugins" / "1c" / "prompts" / "rag-answer.md"
+
+
+def format_context(results: list[dict], *, max_chars: int = 12000) -> str:
+    if not results:
+        return "Контекст не найден."
+
+    blocks = []
+    used_chars = 0
+    for position, result in enumerate(results, start=1):
+        document = result["document"]
+        source = document.get("source_path") or "unknown"
+        chunk = document.get("chunk_index")
+        title = document.get("title") or "unknown"
+        content = (document.get("content") or "").strip()
+        header = f"[{position}] source={source} title={title} chunk={chunk} score={result['score']:.4f}"
+        remaining = max_chars - used_chars - len(header) - 2
+        if remaining <= 0:
+            break
+        if len(content) > remaining:
+            content = content[: max(0, remaining - 3)].rstrip() + "..."
+        block = "\n".join([header, content])
+        blocks.append(block)
+        used_chars += len(block) + 2
+    return "\n\n".join(blocks)
+
+
+def render_prompt(template_path: Path, context: str, question: str) -> str:
+    template = template_path.read_text(encoding="utf-8")
+    return template.replace("{{context}}", context).replace("{{question}}", question)
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description="Ask the 1C RAG assistant.")
+    parser.add_argument("question")
+    parser.add_argument("--index", type=Path, default=DEFAULT_INDEX)
+    parser.add_argument("--profile", default="auto")
+    parser.add_argument("--limit", type=int)
+    parser.add_argument("--candidate-limit", type=int)
+    parser.add_argument("--dedupe-by-document", action="store_true")
+    parser.add_argument("--min-score", type=float)
+    parser.add_argument("--source-type", action="append", dest="source_types")
+    parser.add_argument("--platform-version")
+    parser.add_argument("--platform-doc-id")
+    parser.add_argument("--max-context-chars", type=int)
+    parser.add_argument("--base-url", help="OpenAI-compatible endpoint, for example http://docker-gpu.cin.su:8000")
+    parser.add_argument("--model", default="qwen3-4b-instruct")
+    parser.add_argument("--print-prompt", action="store_true", help="Print assembled prompt instead of calling a model.")
+    parser.add_argument("--system-prompt", type=Path, default=DEFAULT_SYSTEM_PROMPT)
+    parser.add_argument("--rag-prompt", type=Path, default=DEFAULT_RAG_PROMPT)
+    args = parser.parse_args()
+
+    index = read_json(args.index)
+    profile = resolve_rag_profile(args.profile, args.question)
+    source_types = args.source_types if args.source_types is not None else profile["source_types"]
+    results = search_lexical_index(
+        index,
+        args.question,
+        limit=int(args.limit or profile["limit"]),
+        candidate_limit=int(args.candidate_limit or profile["candidate_limit"]),
+        dedupe_by_document=args.dedupe_by_document or bool(profile["dedupe_by_document"]),
+        min_score=float(args.min_score if args.min_score is not None else profile["min_score"]),
+        source_types=source_types,
+        metadata_filters={
+            "platform_version": args.platform_version or "",
+            "platform_doc_id": args.platform_doc_id or "",
+        },
+    )
+    context = format_context(results, max_chars=int(args.max_context_chars or profile["max_context_chars"]))
+    rag_prompt = render_prompt(args.rag_prompt, context=context, question=args.question)
+
+    if args.print_prompt or not args.base_url:
+        print(rag_prompt)
+        if not args.base_url and not args.print_prompt:
+            print(
+                "\nNo --base-url provided, so only the prompt was rendered.",
+                file=sys.stderr,
+            )
+        return 0
+
+    system_prompt = args.system_prompt.read_text(encoding="utf-8")
+    try:
+        answer = call_chat_completion(
+            base_url=args.base_url,
+            model=args.model,
+            messages=[
+                {"role": "system", "content": system_prompt},
+                {"role": "user", "content": rag_prompt},
+            ],
+            max_tokens=1200,
+        )
+    except (urllib.error.URLError, ValueError) as exc:
+        print(f"Chat request failed: {exc}", file=sys.stderr)
+        return 1
+
+    print(answer)
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())