feat: 标签支持多选后直接检索，无关键词时可纯标签检索
- search_result.dspy: 标签解析提前，query为空但标签有值时走直接DB查询 - search.ui: tag_selector添加changed事件触发检索，更新首页提示文案
This commit is contained in:
ymq
parent
70a958180f
commit
3425c09188
@ -3,7 +3,7 @@
|
||||
"options": {"padding": "20px", "spacing": "12px", "width": "100%", "height": "100%", "css": "filler"},
|
||||
"subwidgets": [
|
||||
{"widgettype": "Text", "options": {"text": "🔍 知识检索", "cfontsize": 22, "fontWeight": "bold", "color": "#1a1a2e"}},
|
||||
{"widgettype": "Text", "options": {"text": "选择知识库，输入文字或上传文件检索", "cfontsize": 13, "color": "#888"}},
|
||||
{"widgettype": "Text", "options": {"text": "选择知识库，输入文字、选择标签或上传文件检索", "cfontsize": 13, "color": "#888"}},
|
||||
{
|
||||
"widgettype": "HBox",
|
||||
"options": {"spacing": "12px", "alignItems": "center"},
|
||||
@ -64,6 +64,22 @@
|
||||
}
|
||||
}
|
||||
},
|
||||
{
|
||||
"wid": "tag_selector",
|
||||
"event": "changed",
|
||||
"actiontype": "urlwidget",
|
||||
"target": "search_results",
|
||||
"mode": "replace",
|
||||
"options": {
|
||||
"url": "/rag/knowledge_bases_list/search_result.dspy",
|
||||
"params": {
|
||||
"_webbricks_": 1,
|
||||
"kb_id": "{{kb_selector}}",
|
||||
"keyword": "{{search_bar}}",
|
||||
"tag_ids": "{{tag_selector}}"
|
||||
}
|
||||
}
|
||||
},
|
||||
{
|
||||
"wid": "search_file",
|
||||
"event": "changed",
|
||||
|
||||
@ -34,16 +34,7 @@ if not file_data:
|
||||
import base64, io
|
||||
top_k = int(ns.get('top_k', 5))
|
||||
|
||||
if not query:
|
||||
return json.dumps({
|
||||
"widgettype": "VBox", "options": {"padding": "20px", "spacing": "16px"},
|
||||
"subwidgets": [
|
||||
{"widgettype": "Text", "options": {"text": "🔍 知识检索", "cfontsize": 20, "fontWeight": "bold"}},
|
||||
{"widgettype": "Text", "options": {"text": "请输入检索内容", "color": "#666", "marginTop": "20px"}}
|
||||
]
|
||||
}, ensure_ascii=False)
|
||||
|
||||
# 0. Parse tag filter — find doc_ids that match ALL selected tags
|
||||
# 0. Parse tag filter FIRST — find doc_ids that match ALL selected tags
|
||||
tag_doc_ids = None
|
||||
tag_info = ""
|
||||
if tag_ids_str:
|
||||
@ -65,93 +56,128 @@ if tag_ids_str:
|
||||
"GROUP BY media_id HAVING COUNT(DISTINCT tag_id)=" + str(len(wanted_tags)))
|
||||
recs = await sor.sqlExe(sql, nsq)
|
||||
tag_doc_ids = set(r.media_id for r in recs)
|
||||
tag_info = " (标签过滤: " + ",".join(wanted_tags[:3]) + ("..." if len(wanted_tags)>3 else "") + ")"
|
||||
tag_info = " (标签: " + ",".join(wanted_tags[:3]) + ("..." if len(wanted_tags)>3 else "") + ")"
|
||||
info(f'[search_result] tag filter: {len(tag_doc_ids)} docs match {len(wanted_tags)} tags')
|
||||
except Exception as e:
|
||||
info(f'[search_result] tag filter error: {e}')
|
||||
|
||||
# 1. Embed
|
||||
vec = []
|
||||
try:
|
||||
client = StreamHttpClient()
|
||||
resp = await client.request('POST', 'https://embedding.opencomputing.net:10443/api/embed',
|
||||
json={"texts": [query], "model": "CLIP-ViT-H-14"})
|
||||
emb = json.loads(resp)
|
||||
vec = emb.get("text_embeddings", emb.get("embeddings", [[]]))[0]
|
||||
except:
|
||||
pass
|
||||
if not vec:
|
||||
return json.dumps({"widgettype": "Text", "options": {"text": "向量化失败", "cfontsize": 14, "color": "#e74c3c"}}, ensure_ascii=False)
|
||||
|
||||
# 2. VDB search
|
||||
raw_rows = []
|
||||
try:
|
||||
recall_n = top_k * 3
|
||||
client2 = StreamHttpClient()
|
||||
resp2 = await client2.request('POST', 'https://vectordb.opencomputing.net:10443/v1/query',
|
||||
json={"colname": kb_id, "vector": vec, "pagerows": recall_n, "output_fields": ["*"]})
|
||||
vdb = json.loads(resp2)
|
||||
raw_rows = vdb.get("data", {}).get("rows", [])
|
||||
if not isinstance(raw_rows, list):
|
||||
raw_rows = []
|
||||
except:
|
||||
pass
|
||||
|
||||
# 3. Keyword recall (hybrid search: vector alone misses technical terms)
|
||||
kw_ids = set()
|
||||
kw_rows = []
|
||||
try:
|
||||
tokens = [t for t in query.split() if t][:5] or [query]
|
||||
conds = []
|
||||
nsq = {"kb_id": kb_id}
|
||||
for i, t in enumerate(tokens):
|
||||
conds.append("content LIKE ${kw_" + str(i) + "}$")
|
||||
nsq["kw_" + str(i)] = "%" + t + "%"
|
||||
ksql = "SELECT id, doc_id, content FROM document_chunks WHERE kb_id=${kb_id}$ AND (" + " OR ".join(conds) + ") LIMIT 20"
|
||||
async with get_sor_context(env, 'rag') as sor:
|
||||
krecs = await sor.sqlExe(ksql, nsq)
|
||||
for r in krecs:
|
||||
kw_ids.add(r.id)
|
||||
kw_rows.append({"id": r.id, "doc_id": r.doc_id or "", "score": 0.99, "text": r.content or '', "kw": True})
|
||||
except Exception as e:
|
||||
info('[search_result] keyword recall failed: %s' % e)
|
||||
# Early return only when nothing is provided
|
||||
if not query and not tag_ids_str and not file_data:
|
||||
return json.dumps({
|
||||
"widgettype": "VBox", "options": {"padding": "20px", "spacing": "16px"},
|
||||
"subwidgets": [
|
||||
{"widgettype": "Text", "options": {"text": "🔍 知识检索", "cfontsize": 20, "fontWeight": "bold"}},
|
||||
{"widgettype": "Text", "options": {"text": "请输入检索内容或选择标签", "color": "#666", "marginTop": "20px"}}
|
||||
]
|
||||
}, ensure_ascii=False)
|
||||
|
||||
hits = []
|
||||
seen = set()
|
||||
for row in raw_rows:
|
||||
rid = str(row.get("id", ""))
|
||||
score = row.get("score", 0)
|
||||
# Resolve doc_id from chunk id pattern "xxxx_c0"
|
||||
doc_id = rid.rsplit('_c', 1)[0] if '_c' in rid else rid
|
||||
# Look up chunk text from DB
|
||||
chunk_text = ''
|
||||
async with get_sor_context(env, 'rag') as sor:
|
||||
recs = await sor.sqlExe(
|
||||
"SELECT content, doc_id FROM document_chunks WHERE id=${id}$",
|
||||
{"id": rid})
|
||||
if recs:
|
||||
chunk_text = recs[0].content or ''
|
||||
doc_id = recs[0].doc_id or doc_id
|
||||
# Tag filter
|
||||
if tag_doc_ids is not None and doc_id not in tag_doc_ids:
|
||||
continue
|
||||
if chunk_text:
|
||||
is_kw = rid in kw_ids
|
||||
hits.append({"id": rid, "doc_id": doc_id, "score": 0.99 if is_kw else score, "text": chunk_text, "kw": is_kw})
|
||||
seen.add(rid)
|
||||
for kr in kw_rows:
|
||||
if kr["id"] not in seen:
|
||||
if tag_doc_ids is not None and kr["doc_id"] not in tag_doc_ids:
|
||||
raw_rows = []
|
||||
kw_rows = []
|
||||
|
||||
if query:
|
||||
# 1. Embed
|
||||
vec = []
|
||||
try:
|
||||
client = StreamHttpClient()
|
||||
resp = await client.request('POST', 'https://embedding.opencomputing.net:10443/api/embed',
|
||||
json={"texts": [query], "model": "CLIP-ViT-H-14"})
|
||||
emb = json.loads(resp)
|
||||
vec = emb.get("text_embeddings", emb.get("embeddings", [[]]))[0]
|
||||
except:
|
||||
pass
|
||||
if not vec:
|
||||
return json.dumps({"widgettype": "Text", "options": {"text": "向量化失败", "cfontsize": 14, "color": "#e74c3c"}}, ensure_ascii=False)
|
||||
|
||||
# 2. VDB search
|
||||
try:
|
||||
recall_n = top_k * 3
|
||||
client2 = StreamHttpClient()
|
||||
resp2 = await client2.request('POST', 'https://vectordb.opencomputing.net:10443/v1/query',
|
||||
json={"colname": kb_id, "vector": vec, "pagerows": recall_n, "output_fields": ["*"]})
|
||||
vdb = json.loads(resp2)
|
||||
raw_rows = vdb.get("data", {}).get("rows", [])
|
||||
if not isinstance(raw_rows, list):
|
||||
raw_rows = []
|
||||
except:
|
||||
pass
|
||||
|
||||
# 3. Keyword recall (hybrid search: vector alone misses technical terms)
|
||||
kw_ids = set()
|
||||
try:
|
||||
tokens = [t for t in query.split() if t][:5] or [query]
|
||||
conds = []
|
||||
nsq = {"kb_id": kb_id}
|
||||
for i, t in enumerate(tokens):
|
||||
conds.append("content LIKE ${kw_" + str(i) + "}$")
|
||||
nsq["kw_" + str(i)] = "%" + t + "%"
|
||||
ksql = "SELECT id, doc_id, content FROM document_chunks WHERE kb_id=${kb_id}$ AND (" + " OR ".join(conds) + ") LIMIT 20"
|
||||
async with get_sor_context(env, 'rag') as sor:
|
||||
krecs = await sor.sqlExe(ksql, nsq)
|
||||
for r in krecs:
|
||||
kw_ids.add(r.id)
|
||||
kw_rows.append({"id": r.id, "doc_id": r.doc_id or "", "score": 0.99, "text": r.content or '', "kw": True})
|
||||
except Exception as e:
|
||||
info('[search_result] keyword recall failed: %s' % e)
|
||||
|
||||
seen = set()
|
||||
for row in raw_rows:
|
||||
rid = str(row.get("id", ""))
|
||||
score = row.get("score", 0)
|
||||
# Resolve doc_id from chunk id pattern "xxxx_c0"
|
||||
doc_id = rid.rsplit('_c', 1)[0] if '_c' in rid else rid
|
||||
# Look up chunk text from DB
|
||||
chunk_text = ''
|
||||
async with get_sor_context(env, 'rag') as sor:
|
||||
recs = await sor.sqlExe(
|
||||
"SELECT content, doc_id FROM document_chunks WHERE id=${id}$",
|
||||
{"id": rid})
|
||||
if recs:
|
||||
chunk_text = recs[0].content or ''
|
||||
doc_id = recs[0].doc_id or doc_id
|
||||
# Tag filter
|
||||
if tag_doc_ids is not None and doc_id not in tag_doc_ids:
|
||||
continue
|
||||
hits.append(kr)
|
||||
seen.add(kr["id"])
|
||||
if chunk_text:
|
||||
is_kw = rid in kw_ids
|
||||
hits.append({"id": rid, "doc_id": doc_id, "score": 0.99 if is_kw else score, "text": chunk_text, "kw": is_kw})
|
||||
seen.add(rid)
|
||||
for kr in kw_rows:
|
||||
if kr["id"] not in seen:
|
||||
if tag_doc_ids is not None and kr["doc_id"] not in tag_doc_ids:
|
||||
continue
|
||||
hits.append(kr)
|
||||
seen.add(kr["id"])
|
||||
|
||||
hits.sort(key=lambda x: x.get("score", 0), reverse=True)
|
||||
hits = hits[:top_k]
|
||||
hits.sort(key=lambda x: x.get("score", 0), reverse=True)
|
||||
hits = hits[:top_k]
|
||||
|
||||
else:
|
||||
# Tag-only search: direct DB lookup for chunks of tagged documents
|
||||
if tag_doc_ids:
|
||||
try:
|
||||
async with get_sor_context(env, 'rag') as sor:
|
||||
kb_cond = ""
|
||||
nsq2 = {}
|
||||
if kb_id and kb_id != "all":
|
||||
kb_cond = "kb_id=${kb_id}$ AND "
|
||||
nsq2["kb_id"] = kb_id
|
||||
placeholders2 = []
|
||||
for i, did in enumerate(tag_doc_ids):
|
||||
placeholders2.append("${did_" + str(i) + "}$")
|
||||
nsq2["did_" + str(i)] = did
|
||||
sql2 = ("SELECT id, doc_id, content FROM document_chunks WHERE " + kb_cond +
|
||||
"doc_id IN (" + ",".join(placeholders2) + ") ORDER BY updated_at DESC LIMIT " + str(top_k))
|
||||
crecs = await sor.sqlExe(sql2, nsq2)
|
||||
for r in crecs:
|
||||
hits.append({"id": r.id, "doc_id": r.doc_id or "", "score": 1.0, "text": r.content or '', "kw": False})
|
||||
except Exception as e:
|
||||
info(f'[search_result] tag-only lookup error: {e}')
|
||||
|
||||
header_text = f"🔍 检索: {query}" if query else "🔍 标签检索"
|
||||
subwidgets = [
|
||||
{"widgettype": "Text", "options": {"text": f"🔍 检索: {query}{tag_info}", "cfontsize": 18, "fontWeight": "bold", "marginBottom": "8px"}},
|
||||
{"widgettype": "Text", "options": {"text": f"共 {len(hits)} 条结果 (召回 {len(raw_rows)} 条)", "cfontsize": 13, "color": "#888", "marginBottom": "16px"}}
|
||||
{"widgettype": "Text", "options": {"text": header_text + tag_info, "cfontsize": 18, "fontWeight": "bold", "marginBottom": "8px"}},
|
||||
{"widgettype": "Text", "options": {"text": f"共 {len(hits)} 条结果" + (f" (召回 {len(raw_rows)} 条)" if raw_rows else ""), "cfontsize": 13, "color": "#888", "marginBottom": "16px"}}
|
||||
]
|
||||
|
||||
if not hits:
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user