Skip to content

RAGFlow is an agent-native database: a Retrieval-Augmented Generation (RAG) engine that grew into context infrastructure.

Fix agentic search bugs (#19615)

Yingfeng2477151c65b3
diff --git a/api/db/services/dialog_service.py b/api/db/services/dialog_service.py
index 6ce31c5f8b..27ab24471b 100644
--- a/api/db/services/dialog_service.py
+++ b/api/db/services/dialog_service.py
@@ -1884,13 +1884,38 @@ async def gen_mindmap(question, kb_ids, tenant_id, search_config={}):
return mind_map.output
+def _bound_dataset_names(dialog) -> str:
+ """Comma-joined names of the dialog's bound datasets ("" when none).
+
+ Exposed through the agentic graph's untrusted evidence block (appended to
+ the "Evidence:" section of the compose user content): the prompt must
+ still NAME the bound datasets, but as runtime data they must not be
+ injected into the reasoning system prompt via ``{knowledge}`` — that
+ placeholder is trusted-template content only. Defaulting ``{knowledge}``
+ to the names rendered an empty `` `` `` on first turn (templates like
+ "derived solely from this dataset: ``{knowledge}``"), which the outer
+ model read as "the dataset is empty" — answering the canned "not found
+ in the dataset!" line without ever calling the ``rag`` tool (first-turn
+ short-circuit, observed 2026-09-14).
+ """
+ if not getattr(dialog, "kb_ids", None):
+ return ""
+ try:
+ kbs = KnowledgebaseService.get_by_ids(dialog.kb_ids) or []
+ except Exception: # noqa: BLE001 — names are cosmetic; never block the render
+ return ""
+ return ", ".join(kb.name for kb in kbs if getattr(kb, "name", ""))
+
+
def _render_reasoning_system_prompt(dialog, prompt_config: dict, kwargs: dict) -> str:
"""Render the dialog-level system prompt for the reasoning agent path.
Mirrors the substitutions ``async_chat`` performs for the non-reasoning path
so that configured system prompts are honored when reasoning is enabled.
- The ``{knowledge}`` placeholder is defaulted to an empty string because the
- agentic graph supplies retrieved evidence through its own evidence block.
+ The ``{knowledge}`` placeholder is trusted-template content only: when the
+ caller does not supply a value it renders empty. Bound dataset names are
+ runtime data — they are exposed through the agentic graph's untrusted
+ evidence block, never through the system prompt.
"""
system = prompt_config.get("system", "")
if not system:
@@ -1902,7 +1927,6 @@ def _render_reasoning_system_prompt(dialog, prompt_config: dict, kwargs: dict) -
param_keys = [p["key"] for p in prompt_config.get("parameters", [])]
if dialog.kb_ids and "knowledge" not in param_keys and "{knowledge}" in system:
param_keys.append("knowledge")
- kwargs.setdefault("knowledge", "")
for p in prompt_config.get("parameters", []):
if p["key"] == "knowledge":
@@ -1913,6 +1937,11 @@ def _render_reasoning_system_prompt(dialog, prompt_config: dict, kwargs: dict) -
system = system.replace("{%s}" % p["key"], " ")
fmt_kwargs = dict(kwargs)
+ # {knowledge} is trusted-template content only: mutable runtime data
+ # (bound dataset names) must never be injected into the system prompt
+ # through it. The bound dataset names are exposed through the untrusted
+ # evidence block instead (see the agentic compose in
+ # rag/advanced_rag/agentic_rag_graph.py).
fmt_kwargs.setdefault("knowledge", "")
return system.format(**fmt_kwargs)
@@ -2088,6 +2117,9 @@ async def rag_agent(dialog, messages, stream=True, **kwargs):
rerank_candidates_count=dialog.rerank_candidates_count,
top_k=dialog.top_k,
)
+ # Mutable runtime data: exposed via the untrusted evidence block in the
+ # agentic compose, never via the reasoning system prompt's {knowledge}.
+ rag_tools._bound_dataset_names = _bound_dataset_names(dialog)
async def decorate_answer(answer):
nonlocal rag_tools, messages
diff --git a/cmd/ragflow_server.go b/cmd/ragflow_server.go
index 310b557d9a..0e085d38d6 100644
--- a/cmd/ragflow_server.go
+++ b/cmd/ragflow_server.go
@@ -1054,7 +1054,7 @@ func startServer(ctx context.Context) {
Images: req.Images,
TextAttachments: req.TextAttachments,
})
- res := service.HarnessResult{Chunks: r.Chunks, DocAggs: r.DocAggs, Answer: r.Answer}
+ res := service.HarnessResult{Chunks: r.Chunks, DocAggs: r.DocAggs, Answer: r.Answer, SlotCitations: r.SlotCitations}
if r.Kbinfos != nil {
res.Memory = r.Kbinfos.Memory
res.PreSummary = r.Kbinfos.PreSummary
diff --git a/internal/rag/advanced_rag/agentic_rag.go b/internal/rag/advanced_rag/agentic_rag.go
index 5f08d4284f..79966e495c 100644
--- a/internal/rag/advanced_rag/agentic_rag.go
+++ b/internal/rag/advanced_rag/agentic_rag.go
@@ -572,6 +572,12 @@ type RunResponse struct {
// Kbinfos carries the full accumulated state (including the lossless
// memory store) for callers that need more than the summary above.
Kbinfos *harness.Kbinfos
+ // SlotCitations maps a slot-table id ("0", "1", ...) to the evidence
+ // chunk ids that filled the slot. The chat pipeline's citation decoration
+ // uses it to rewrite leaked "[ID:Slot N]" markers into citations of the
+ // chunk the slot was filled from (those markers index the internal slot
+ // table — nothing the user can open).
+ SlotCitations map[string][]string
}
// AnswerSink forwards a partially produced answer while the model is still
diff --git a/internal/rag/advanced_rag/agentic_rag_graph.go b/internal/rag/advanced_rag/agentic_rag_graph.go
index 0c49d3854f..f8ebd75c1f 100644
--- a/internal/rag/advanced_rag/agentic_rag_graph.go
+++ b/internal/rag/advanced_rag/agentic_rag_graph.go
@@ -3548,6 +3548,17 @@ func NewAgenticLoop() AgenticLoop {
// Surface the loop's outcome on the RunResponse.
resp.Slots = append(resp.Slots, st.SlotTable.State...)
+ // Slot citations: the chat pipeline's citation decoration rewrites
+ // leaked "[ID:Slot N]" markers with these evidence chunk ids. Plain
+ // ids (not positions) so the mapping survives the caller's later
+ // evidence-pool narrowing (selectEvidence).
+ if len(st.SlotEvidence) > 0 {
+ sc := make(map[string][]string, len(st.SlotEvidence))
+ for sid, ev := range st.SlotEvidence {
+ sc[sid] = ev.EvidenceIDs
+ }
+ resp.SlotCitations = sc
+ }
// Research findings: the SCA-reviewed draft. NOTE: this is the research
// draft, NOT the final answer — RAGTools.Run composes the final cited
// answer afterwards from KB.PreSummary (see composeFinalAnswer), exactly
diff --git a/internal/rag/prompts/citation_prompt.md b/internal/rag/prompts/citation_prompt.md
index 3db2814289..ca9b208a78 100644
--- a/internal/rag/prompts/citation_prompt.md
+++ b/internal/rag/prompts/citation_prompt.md
@@ -4,6 +4,7 @@ Based on the provided document or chat history, add citations to the input text
## Technical Rules:
- Use format: [ID:i] or [ID:i] [ID:j] for multiple sources
+- Cite each source INDIVIDUALLY: [ID:i][ID:j]. NEVER merge consecutive citations into a range such as [ID:i-j] or [ID:i~j]
- Place citations at the end of sentences, before punctuation
- Maximum 4 citations per sentence
- DO NOT cite content not from <context></context>
@@ -120,3 +121,4 @@ REMEMBER:
- When in doubt, ask: "Would a fact-checker need to verify this?"
- Place citations at sentence end, before punctuation
- Format likes this is FORBIDDEN: [ID:0, ID:5, ID:...]. It MUST be separated like, [ID:0][ID:5]...
+- Range citations like [ID:1-3] are FORBIDDEN too. Multiple sources MUST be spelled out individually: [ID:1][ID:2][ID:3].
diff --git a/internal/rag/prompts/generator.go b/internal/rag/prompts/generator.go
index 8078dda229..d60c5caae0 100644
--- a/internal/rag/prompts/generator.go
+++ b/internal/rag/prompts/generator.go
@@ -113,7 +113,8 @@ var DefaultCiteRules = CitationPrompt("")
// appends after rendering CITATION_PROMPT_TEMPLATE: the examples in the
// template carry illustrative IDs, and without this the model can emit them
// instead of the real chunk IDs it was given.
-const citationIDSuffix = "\n\nIMPORTANT: The example IDs above (45, 46, 78, etc.) are illustrative only. Use the actual chunk IDs from the provided knowledge blocks."
+const citationIDSuffix = "\n\nIMPORTANT: The example IDs above (45, 46, 78, etc.) are illustrative only. Use the actual chunk IDs from the provided knowledge blocks." +
+ " Cite each source INDIVIDUALLY as [ID:i][ID:j]; NEVER merge consecutive citations into a range such as [ID:i-j]."
// CitationPrompt mirrors Python rag/prompts/generator.citation_prompt: return
// the citation-rules text the final answer must follow. The user-defined
diff --git a/internal/service/chat_harness_prompt_test.go b/internal/service/chat_harness_prompt_test.go
new file mode 100644
index 0000000000..df9ca93bbd
--- /dev/null
+++ b/internal/service/chat_harness_prompt_test.go
@@ -0,0 +1,91 @@
+package service
+
+import (
+ "strings"
+ "testing"
+
+ "ragflow/internal/entity"
+)
+
+// Regression lock for the first-turn empty dataset binding: the agentic
+// (reasoning) system prompt used to render {knowledge} as "" — the web UI's
+// default template then read "derived solely from this dataset: ``" and the
+// outer model answered the canned "not found in the dataset!" line without
+// calling the terminal `rag` tool (logs/api_server_9384.log 2026-09-14
+// 13:50:14: answer_chars=59, zero graph LLM calls).
+
+func TestHarnessBoundDatasetNames(t *testing.T) {
+ cases := []struct {
+ name string
+ kbs []*entity.Knowledgebase
+ want string
+ }{
+ {"nil", nil, ""},
+ {"empty", []*entity.Knowledgebase{}, ""},
+ {
+ "filters nil and unnamed",
+ []*entity.Knowledgebase{nil, {ID: "kb1"}, {ID: "kb2", Name: "三国演义"}},
+ "三国演义",
+ },
+ {
+ "joins multiple names",
+ []*entity.Knowledgebase{{Name: "三国演义"}, {Name: "三国志"}},
+ "三国演义, 三国志",
+ },
+ }
+ for _, c := range cases {
+ t.Run(c.name, func(t *testing.T) {
+ if got := harnessBoundDatasetNames(c.kbs); got != c.want {
+ t.Fatalf("harnessBoundDatasetNames(%v) = %q, want %q", c.kbs, got, c.want)
+ }
+ })
+ }
+}
+
+// TestHarnessSystemPromptFirstTurnNonEmpty pins the end-to-end first-turn
+// render: with datasets bound, the dialog system prompt's {knowledge}
+// placeholder (backtick-wrapped in the web UI default) must carry the bound
+// dataset names — never an empty “ binding.
+func TestHarnessSystemPromptFirstTurnNonEmpty(t *testing.T) {
+ // The web UI default template (web/src/locales en.ts systemInitialValue):
+ // the placeholder is wrapped in backticks.
+ template := "You are an intelligent assistant. Your primary function is to answer questions based strictly on the provided knowledge base.\n\n**Essential Rules:**\n - Your answer must be derived **solely** from this dataset: `{knowledge}`.\n"
+
+ s := &ChatPipelineService{}
+ kws := map[string]interface{}{
+ "date": "2026-09-14 06:00:00",
+ }
+ if _, ok := kws["knowledge"]; !ok {
+ kws["knowledge"] = harnessBoundDatasetNames([]*entity.Knowledgebase{
+ {ID: "kb1", Name: "三国演义"},
+ })
+ }
+ rendered := s.formatPrompt(template, kws)
+
+ if strings.Contains(rendered, "``") {
+ t.Fatalf("first-turn system prompt rendered an empty dataset binding:\n%s", rendered)
+ }
+ if !strings.Contains(rendered, "三国演义") {
+ t.Fatalf("first-turn system prompt does not carry the bound dataset name:\n%s", rendered)
+ }
+ if strings.Contains(rendered, "{knowledge}") {
+ t.Fatalf("{knowledge} placeholder was not substituted:\n%s", rendered)
+ }
+}
+
+// TestHarnessSystemPromptCallerSuppliedKnowledgeWins: a caller-supplied
+// knowledge value must not be overwritten by the dataset-name default.
+func TestHarnessSystemPromptCallerSuppliedKnowledgeWins(t *testing.T) {
+ s := &ChatPipelineService{}
+ kws := map[string]interface{}{"knowledge": "caller evidence"}
+ if _, ok := kws["knowledge"]; !ok {
+ kws["knowledge"] = harnessBoundDatasetNames([]*entity.Knowledgebase{{Name: "三国演义"}})
+ }
+ rendered := s.formatPrompt("Context: {knowledge}", kws)
+ if !strings.Contains(rendered, "caller evidence") {
+ t.Fatalf("caller-supplied knowledge was overwritten:\n%s", rendered)
+ }
+ if strings.Contains(rendered, "三国演义") {
+ t.Fatalf("dataset-name default leaked over the caller-supplied value:\n%s", rendered)
+ }
+}
diff --git a/internal/service/chat_pipeline.go b/internal/service/chat_pipeline.go
index ba67310463..e7661b9f3a 100644
--- a/internal/service/chat_pipeline.go
+++ b/internal/service/chat_pipeline.go
@@ -743,7 +743,12 @@ func (s *ChatPipelineService) AsyncChat(
// dialog, prompt_config, kwargs) to RAGTools
// (dialog_service.py:2084): the dialog system prompt rendered
// with the caller kwargs, a UTC date, and {knowledge} defaulted
- // to "" — the agentic graph supplies evidence itself.
+ // to the BOUND DATASET NAMES — the agentic graph supplies the
+ // evidence itself, but an empty binding made the outer model
+ // read the prompt as "the dataset is empty" and answer the
+ // canned "not found in the dataset!" line without calling the
+ // terminal `rag` tool (first-turn short-circuit, observed
+ // 2026-09-14: answer_chars=59, zero graph LLM calls).
harnessSystemPrompt := ""
if sp, ok := chat.PromptConfig["system"].(string); ok && sp != "" {
kws := make(map[string]interface{}, len(kwargs)+2)
@@ -752,11 +757,11 @@ func (s *ChatPipelineService) AsyncChat(
}
kws["date"] = time.Now().UTC().Format("2006-01-02 15:04:05")
if _, ok := kws["knowledge"]; !ok {
- kws["knowledge"] = ""
+ kws["knowledge"] = harnessBoundDatasetNames(kbs)
}
harnessSystemPrompt = s.formatPrompt(sp, kws)
}
- hk, harnessAnswer, hErr := s.retrieveViaHarness(ctx, question, kbs, docIDs, imageFiles, attachments, thinkingMode, chat.TenantID, chat.LLMID, chat.ID, webSearch, sink, harnessSystemPrompt)
+ hk, slotCites, harnessAnswer, hErr := s.retrieveViaHarness(ctx, question, kbs, docIDs, imageFiles, attachments, thinkingMode, chat.TenantID, chat.LLMID, chat.ID, webSearch, sink, harnessSystemPrompt)
// The harness streams think-then-answer inside ONE compose call.
// Close the block here, once that call (and its trailing
// narration line) has returned: a reasoning-only run would
@@ -771,7 +776,7 @@ func (s *ChatPipelineService) AsyncChat(
if harnessAnswer != "" {
common.Info("harness produced final cited answer; short-circuiting",
zap.Int("answer_chars", len(harnessAnswer)))
- final := s.decorateHarnessAnswer(harnessAnswer, kbinfos)
+ final := s.decorateHarnessAnswer(harnessAnswer, kbinfos, slotCites)
final.Final = true
out <- final
return
@@ -3079,7 +3084,7 @@ func (s *ChatPipelineService) decorateAnswer(
// here. We only resolve the existing markers, repair bad formats, filter
// doc_aggs to the cited docs, and build the reference from the harness citation
// pool (chunks stripped of their vectors). Prompt stays empty, matching Python.
-func (s *ChatPipelineService) decorateHarnessAnswer(answer string, kbinfos map[string]interface{}) AsyncChatResult {
+func (s *ChatPipelineService) decorateHarnessAnswer(answer string, kbinfos map[string]interface{}, slotCitations map[string][]string) AsyncChatResult {
think := ""
ans := answer
if strings.Contains(answer, "</think>") {
@@ -3091,6 +3096,18 @@ func (s *ChatPipelineService) decorateHarnessAnswer(answer string, kbinfos map[s
chunksRaw, _ := kbinfos["chunks"].([]map[string]interface{})
+ // Slot-table markers ("[ID:Slot 0]") are internal references the compose
+ // model leaks from the Research Summary's slot draft — they index the slot
+ // table, not any chunk the user can open. Rewrite them into citations of
+ // the chunk that filled the slot (or drop them) BEFORE the marker scan, so
+ // they resolve like any other citation.
+ ans = RepairSlotCitations(ans, slotCitations, chunksRaw)
+
+ // Range-merged citations ("[ID:1-3]") are the model compressing
+ // consecutive individual citations on its own; expand them back so every
+ // marker resolves to exactly one chunk (out-of-range ranges are dropped).
+ ans = ExpandRangeCitations(ans, len(chunksRaw))
+
// Collect existing [ID:N] markers from the harness answer. Python runs
// normalize_arabic_digits then CITATION_MARKER_PATTERN (dialog_service.py:
// 2103-2107) and bounds each index by len(chunks).
@@ -3197,7 +3214,8 @@ func (s *ChatPipelineService) extractVisibleAnswer(text string) string {
// Mirrors Python's citation_prompt() in rag/prompts/generator.py.
func citationPrompt() string {
return "\n\n### Citation\nWhen answering, please cite sources using the format [ID:N] " +
- "(where N is the chunk number) after each sentence where the information from that chunk is used."
+ "(where N is the chunk number) after each sentence where the information from that chunk is used. " +
+ "Cite each source individually ([ID:1][ID:2]); never merge consecutive citations into a range such as [ID:1-3]."
}
// -----------------------------------------------------------------------
@@ -4657,6 +4675,25 @@ func getChunkValue(chunk map[string]interface{}, k1, k2 string) interface{} {
return chunk[k2]
}
+// harnessBoundDatasetNames renders the {knowledge} default for the agentic
+// (reasoning) system prompt: the comma-joined names of the chat's bound
+// datasets. Python parity: dialog_service._render_reasoning_system_prompt /
+// _bound_dataset_names. The agentic graph supplies the retrieved evidence
+// itself, but the prompt must still NAME the bound datasets — defaulting the
+// placeholder to "" left the web UI's default template rendering as
+// "derived solely from this dataset: “", which the outer model read as an
+// empty dataset and answered the canned "not found in the dataset!" line
+// without ever calling the terminal `rag` tool.
+func harnessBoundDatasetNames(kbs []*entity.Knowledgebase) string {
+ names := make([]string, 0, len(kbs))
+ for _, kb := range kbs {
+ if kb != nil && kb.Name != "" {
+ names = append(names, kb.Name)
+ }
+ }
+ return strings.Join(names, ", ")
+}
+
// HarnessRequest carries the minimal inputs the chat pipeline hands to the
// agentic-RAG harness for evidence collection.
type HarnessRequest struct {
@@ -4714,6 +4751,10 @@ type HarnessResult struct {
// it directly as the reply instead of re-generating via a second model pass —
// mirroring Python's terminal `rag` tool.
Answer string
+ // SlotCitations maps a slot-table id ("0", ...) to the evidence chunk ids
+ // that filled it (RunResponse.SlotCitations). The citation decoration uses
+ // it to rewrite leaked "[ID:Slot N]" markers into locatable chunk citations.
+ SlotCitations map[string][]string
}
// harnessRetriever is wired at server bootstrap (cmd/ragflow_server.go:889) to
@@ -4766,9 +4807,9 @@ func toolLoopLine(sink func(delta string, isThink bool), line string) {
sink(line+thinkLineBreak, true)
}
-func (s *ChatPipelineService) retrieveViaHarness(ctx context.Context, question string, kbs []*entity.Knowledgebase, docIDs []string, images []string, textAttachments string, thinkingMode, tenantID, modelID, sessionID string, webSearch func(context.Context, []string) ([]string, error), answerSink func(delta string, isThink bool), dialogSystemPrompt string) (map[string]interface{}, string, error) {
+func (s *ChatPipelineService) retrieveViaHarness(ctx context.Context, question string, kbs []*entity.Knowledgebase, docIDs []string, images []string, textAttachments string, thinkingMode, tenantID, modelID, sessionID string, webSearch func(context.Context, []string) ([]string, error), answerSink func(delta string, isThink bool), dialogSystemPrompt string) (map[string]interface{}, map[string][]string, string, error) {
if harnessRetriever == nil {
- return nil, "", fmt.Errorf("harness retriever not wired at bootstrap")
+ return nil, nil, "", fmt.Errorf("harness retriever not wired at bootstrap")
}
kbIDs := kbIDStrings(kbs)
// Only non-naive modes run the agentic loop these lines describe, and only
@@ -4793,7 +4834,7 @@ func (s *ChatPipelineService) retrieveViaHarness(ctx context.Context, question s
SystemPrompt: dialogSystemPrompt,
})
if err != nil {
- return nil, "", err
+ return nil, nil, "", err
}
if narrate {
// Python distinguishes the terminal tool from an observation: an answer
@@ -4816,7 +4857,7 @@ func (s *ChatPipelineService) retrieveViaHarness(ctx context.Context, question s
if res.PreSummary != "" {
kbinfos["pre_summary"] = res.PreSummary
}
- return kbinfos, res.Answer, nil
+ return kbinfos, res.SlotCitations, res.Answer, nil
}
// toAnySlice widens a []map[string]any to []interface{} so the doc_aggs field
diff --git a/internal/service/chat_pipeline_test.go b/internal/service/chat_pipeline_test.go
index aa167dd8d2..3fbbc55f11 100644
--- a/internal/service/chat_pipeline_test.go
+++ b/internal/service/chat_pipeline_test.go
@@ -1569,7 +1569,7 @@ func TestRetrieveViaHarnessEmitsToolLoopLines(t *testing.T) {
var got []string
var thinks []bool
s := &ChatPipelineService{}
- _, answer, err := s.retrieveViaHarness(context.Background(), "q", nil, nil, nil, "", "high", "t", "m", "sess", nil, collectSink(&got, &thinks), "")
+ _, _, answer, err := s.retrieveViaHarness(context.Background(), "q", nil, nil, nil, "", "high", "t", "m", "sess", nil, collectSink(&got, &thinks), "")
if err != nil {
t.Fatalf("retrieveViaHarness: %v", err)
}
@@ -1611,7 +1611,7 @@ func TestRetrieveViaHarnessToolLoopObservation(t *testing.T) {
var got []string
var thinks []bool
s := &ChatPipelineService{}
- if _, _, err := s.retrieveViaHarness(context.Background(), "q", nil, nil, nil, "", "high", "t", "m", "sess", nil, collectSink(&got, &thinks), ""); err != nil {
+ if _, _, _, err := s.retrieveViaHarness(context.Background(), "q", nil, nil, nil, "", "high", "t", "m", "sess", nil, collectSink(&got, &thinks), ""); err != nil {
t.Fatalf("retrieveViaHarness: %v", err)
}
joined := strings.Join(got, "")
@@ -1631,7 +1631,7 @@ func TestRetrieveViaHarnessNaiveSkipsToolLoop(t *testing.T) {
var got []string
var thinks []bool
s := &ChatPipelineService{}
- if _, _, err := s.retrieveViaHarness(context.Background(), "q", nil, nil, nil, "", "naive", "t", "m", "sess", nil, collectSink(&got, &thinks), ""); err != nil {
+ if _, _, _, err := s.retrieveViaHarness(context.Background(), "q", nil, nil, nil, "", "naive", "t", "m", "sess", nil, collectSink(&got, &thinks), ""); err != nil {
t.Fatalf("retrieveViaHarness: %v", err)
}
if len(got) != 0 {
@@ -1663,7 +1663,7 @@ func TestDecorateHarnessAnswerReferenceUsesClientChunkShape(t *testing.T) {
}
s := &ChatPipelineService{}
- res := s.decorateHarnessAnswer("The answer [ID:0]", kbinfos)
+ res := s.decorateHarnessAnswer("The answer [ID:0]", kbinfos, nil)
if res.Reference == nil {
t.Fatal("a cited answer must carry a reference")
}
@@ -1683,3 +1683,43 @@ func TestDecorateHarnessAnswerReferenceUsesClientChunkShape(t *testing.T) {
t.Error("the shared source chunk lost its vector: the reference must not mutate it")
}
}
+
+// TestDecorateHarnessAnswerRewritesSlotCitations is the end-to-end lock for
+// the leaked "[ID:Slot 0]" bug: the harness answer arrives citing the internal
+// slot table, and the final answer must cite the chunk that filled the slot
+// instead — with that chunk present in the reference payload.
+func TestDecorateHarnessAnswerRewritesSlotCitations(t *testing.T) {
+ kbinfos := map[string]interface{}{
+ "chunks": []map[string]interface{}{
+ {"chunk_id": "c1", "content_with_weight": "unrelated", "doc_id": "d0", "docnm_kwd": "Other"},
+ {"chunk_id": "c9", "content_with_weight": "eiffel", "doc_id": "d1", "docnm_kwd": "Doc One"},
+ },
+ "doc_aggs": []interface{}{
+ map[string]interface{}{"doc_id": "d1", "doc_name": "Doc One"},
+ },
+ }
+ s := &ChatPipelineService{}
+ res := s.decorateHarnessAnswer(
+ "The tower opened in 1889 [ID:Slot 0].",
+ kbinfos,
+ map[string][]string{"0": {"c9"}},
+ )
+ if strings.Contains(res.Answer, "[ID:Slot") {
+ t.Fatalf("final answer still carries the internal slot citation: %q", res.Answer)
+ }
+ if !strings.Contains(res.Answer, "[ID:1]") {
+ t.Fatalf("final answer must cite the evidence chunk pool position, got %q", res.Answer)
+ }
+ chunks, _ := res.Reference["chunks"].([]map[string]interface{})
+ // The reference keeps the whole citation pool (Python parity); it must at
+ // least carry the slot's evidence chunk so [ID:1] is openable.
+ found := false
+ for _, c := range chunks {
+ if c["content"] == "eiffel" {
+ found = true
+ }
+ }
+ if !found {
+ t.Fatalf("reference must carry the slot's evidence chunk, got %#v", res.Reference["chunks"])
+ }
+}
diff --git a/internal/service/citation.go b/internal/service/citation.go
index d427070b74..eec36c7700 100644
--- a/internal/service/citation.go
+++ b/internal/service/citation.go
@@ -20,6 +20,7 @@ import (
"context"
"math"
"regexp"
+ "strconv"
"strings"
)
@@ -367,3 +368,103 @@ func RepairBadCitationFormats(answer string) string {
}
return working
}
+
+// slotCitationPattern matches slot-table citations that leak into the final
+// answer. The compose prompt's Research Summary renders the slot draft
+// ("- slot 0 [entity]: ..."), and models occasionally cite those internal
+// lines with the citation format instead of the evidence blocks:
+// "[ID:Slot 0]", "[ID: slot 0]", "[Slot 0]". Such markers index the internal
+// slot table — no chunk the user can open — so they must be rewritten into
+// citations of the chunk the slot was filled from, or dropped.
+var slotCitationPattern = regexp.MustCompile(`(?i)\[\s*(?:ID\s*[:: ]*\s*)?slot\s*([0-9]+)\s*\]`)
+
+// RepairSlotCitations rewrites leaked slot-table citations into locatable
+// chunk citations. slotCitations maps a slot-table id ("0", "1", ...) to the
+// evidence chunk ids that produced the slot's candidate; chunks is the
+// citation pool the downstream decoration indexes ("[ID:N]" where N is a
+// 0-based position there). A marker whose slot has no evidence chunk in the
+// pool is dropped outright — an internal marker the user cannot resolve must
+// not reach the answer. Python parity note: the same leak exists upstream
+// (agentic_rag.py:885 only repairs "(ID:5)" shapes; the pre_summary slot
+// draft reaches Python's compose prompt too), so this repair is a Go-side
+// decoration-layer fix under the "every citation must locate a chunk" rule.
+func RepairSlotCitations(answer string, slotCitations map[string][]string, chunks []map[string]interface{}) string {
+ if answer == "" || len(slotCitations) == 0 || !slotCitationPattern.MatchString(answer) {
+ return answer
+ }
+ posByID := make(map[string]int, len(chunks))
+ for i, c := range chunks {
+ if id := chunkCitationID(c); id != "" {
+ if _, dup := posByID[id]; !dup {
+ posByID[id] = i
+ }
+ }
+ }
+ var b strings.Builder
+ b.Grow(len(answer))
+ last := 0
+ for _, m := range slotCitationPattern.FindAllStringSubmatchIndex(answer, -1) {
+ slotID := answer[m[2]:m[3]]
+ repl := ""
+ for _, eid := range slotCitations[slotID] {
+ if pos, ok := posByID[eid]; ok {
+ repl = "[ID:" + strconv.Itoa(pos) + "]"
+ break
+ }
+ }
+ b.WriteString(answer[last:m[0]])
+ b.WriteString(repl)
+ last = m[1]
+ }
+ b.WriteString(answer[last:])
+ return b.String()
+}
+
+// chunkCitationID returns a pool chunk's id under the keys the harness uses
+// (chunk_id, then id — harness.ChunkIDOf's order).
+func chunkCitationID(c map[string]interface{}) string {
+ for _, k := range []string{"chunk_id", "id"} {
+ if s, ok := c[k].(string); ok && s != "" {
+ return s
+ }
+ }
+ return ""
+}
+
+// rangeCitationPattern matches range-merged citations: "[ID:1-3]", "[ID: 1 - 3]",
+// "[ID:1~3]". Models sometimes compress consecutive individual citations
+// ([ID:1][ID:2][ID:3]) into such a range on their own — no code path produces
+// it, and no chunk resolves through it, so it must be normalized defensively.
+var rangeCitationPattern = regexp.MustCompile(`(?i)\[\s*ID\s*[:: ]*\s*([0-9]+)\s*[-–—~~]\s*([0-9]+)\s*\]`)
+
+// ExpandRangeCitations expands range-merged citations back into individual
+// ones: [ID:1-3] → [ID:1][ID:2][ID:3]. A range expands only when BOTH bounds
+// name valid citation-pool positions (0-based, bounded by poolSize); reversed
+// bounds are swapped first. Anything out of range is dropped outright — an
+// unresolvable marker must not reach the user's answer, matching the
+// slot-citation repair above. poolSize is len(kbinfos chunks): the same space
+// the downstream [ID:N] resolution indexes.
+func ExpandRangeCitations(answer string, poolSize int) string {
+ if answer == "" || !rangeCitationPattern.MatchString(answer) {
+ return answer
+ }
+ var b strings.Builder
+ b.Grow(len(answer) * 2)
+ last := 0
+ for _, m := range rangeCitationPattern.FindAllStringSubmatchIndex(answer, -1) {
+ a, _ := strconv.Atoi(normalizeArabicDigits(answer[m[2]:m[3]]))
+ c, _ := strconv.Atoi(normalizeArabicDigits(answer[m[4]:m[5]]))
+ if a > c {
+ a, c = c, a
+ }
+ b.WriteString(answer[last:m[0]])
+ if a >= 0 && c < poolSize {
+ for i := a; i <= c; i++ {
+ b.WriteString("[ID:" + strconv.Itoa(i) + "]")
+ }
+ }
+ last = m[1]
+ }
+ b.WriteString(answer[last:])
+ return b.String()
+}
diff --git a/internal/service/citation_test.go b/internal/service/citation_test.go
index 6d9e446a08..2c03461b75 100644
--- a/internal/service/citation_test.go
+++ b/internal/service/citation_test.go
@@ -333,3 +333,179 @@ func TestRepairBadCitationFormats(t *testing.T) {
}
}
}
+
+// -----------------------------------------------------------------------
+// RepairSlotCitations — leaked "[ID:Slot N]" markers (Bug: slot-table
+// citations in the final answer, unlocatable by the user).
+// -----------------------------------------------------------------------
+
+func TestRepairSlotCitations_MapsSlotToEvidenceChunk(t *testing.T) {
+ chunks := []map[string]interface{}{
+ {"chunk_id": "c1", "content": "one"},
+ {"chunk_id": "c2", "content": "two"},
+ {"id": "c9", "content": "slot evidence"},
+ }
+ answer := "Gustave Eiffel designed it [ID:Slot 0]. It opened in 1889."
+ got := RepairSlotCitations(answer, map[string][]string{"0": {"c9", "missing"}}, chunks)
+ if strings.Contains(got, "Slot") {
+ t.Fatalf("slot marker leaked: %q", got)
+ }
+ if !strings.Contains(got, "[ID:2]") {
+ t.Fatalf("slot marker not mapped to the evidence chunk position: %q", got)
+ }
+ if got != "Gustave Eiffel designed it [ID:2]. It opened in 1889." {
+ t.Fatalf("unexpected rewrite: %q", got)
+ }
+}
+
+func TestRepairSlotCitations_DropsUnresolvable(t *testing.T) {
+ chunks := []map[string]interface{}{{"chunk_id": "c1", "content": "one"}}
+ answer := "Fact one [ID:Slot 0]. Fact two [Slot 1]."
+ got := RepairSlotCitations(answer, map[string][]string{"0": {"not-in-pool"}, "1": {"also-missing"}}, chunks)
+ if strings.Contains(strings.ToLower(got), "slot") {
+ t.Fatalf("unresolvable slot markers must be dropped, got %q", got)
+ }
+ if got != "Fact one . Fact two ." {
+ t.Fatalf("unexpected rewrite: %q", got)
+ }
+}
+
+func TestRepairSlotCitations_MarkerVariants(t *testing.T) {
+ chunks := []map[string]interface{}{{"chunk_id": "a"}, {"id": "b"}}
+ cases := []struct {
+ marker string
+ want string
+ }{
+ {"[ID:Slot 1]", "[ID:1]"},
+ {"[ID: slot 1]", "[ID:1]"},
+ {"[Slot 1]", "[ID:1]"},
+ {"[slot 1]", "[ID:1]"},
+ }
+ for _, c := range cases {
+ got := RepairSlotCitations("x "+c.marker+" y", map[string][]string{"1": {"b"}}, chunks)
+ if !strings.Contains(got, c.want) {
+ t.Errorf("marker %q: got %q, want it rewritten to %q", c.marker, got, c.want)
+ }
+ }
+}
+
+func TestRepairSlotCitations_NoopWithoutSlotMarkersOrMap(t *testing.T) {
+ answer := "Plain answer [ID:0]."
+ if got := RepairSlotCitations(answer, nil, nil); got != answer {
+ t.Fatalf("nil map must be a no-op, got %q", got)
+ }
+ if got := RepairSlotCitations(answer, map[string][]string{"0": {"c"}}, nil); got != answer {
+ t.Fatalf("no slot markers must be a no-op, got %q", got)
+ }
+}
+
+// -----------------------------------------------------------------------
+// ExpandRangeCitations — range-merged citations ("[ID:1-3]") the model
+// produces on its own (no code path emits them).
+// -----------------------------------------------------------------------
+
+func TestExpandRangeCitations(t *testing.T) {
+ cases := []struct {
+ name string
+ answer string
+ poolSize int
+ want string
+ }{
+ {
+ "plain range",
+ "Fact [ID:1-3].",
+ 5,
+ "Fact [ID:1][ID:2][ID:3].",
+ },
+ {
+ "spaced dash variant",
+ "Fact [ID: 1 - 3].",
+ 5,
+ "Fact [ID:1][ID:2][ID:3].",
+ },
+ {
+ "tilde variant",
+ "Fact [ID:1~3].",
+ 5,
+ "Fact [ID:1][ID:2][ID:3].",
+ },
+ {
+ "reversed bounds swap",
+ "Fact [ID:3-1].",
+ 5,
+ "Fact [ID:1][ID:2][ID:3].",
+ },
+ {
+ "single-bound range",
+ "Fact [ID:2-2].",
+ 5,
+ "Fact [ID:2].",
+ },
+ {
+ "out of bounds dropped",
+ "Fact [ID:99-102].",
+ 5,
+ "Fact .",
+ },
+ {
+ "partially out of bounds dropped",
+ "Fact [ID:2-99].",
+ 5,
+ "Fact .",
+ },
+ {
+ "empty pool dropped",
+ "Fact [ID:1-3].",
+ 0,
+ "Fact .",
+ },
+ {
+ "no ranges untouched",
+ "Fact [ID:1] and [ID:2].",
+ 5,
+ "Fact [ID:1] and [ID:2].",
+ },
+ }
+ for _, c := range cases {
+ t.Run(c.name, func(t *testing.T) {
+ if got := ExpandRangeCitations(c.answer, c.poolSize); got != c.want {
+ t.Fatalf("ExpandRangeCitations(%q, %d) = %q, want %q", c.answer, c.poolSize, got, c.want)
+ }
+ })
+ }
+}
+
+// TestDecorateHarnessAnswerExpandsRangeCitations pins the end-to-end fix for
+// the [ID:1-3] report: the final answer carries only individual citations,
+// each resolving to a pooled chunk, and only the cited doc reaches the
+// reference.
+func TestDecorateHarnessAnswerExpandsRangeCitations(t *testing.T) {
+ kbinfos := map[string]interface{}{
+ "chunks": []map[string]interface{}{
+ {"chunk_id": "c0", "content_with_weight": "a", "doc_id": "d1", "docnm_kwd": "Doc One"},
+ {"chunk_id": "c1", "content_with_weight": "b", "doc_id": "d1", "docnm_kwd": "Doc One"},
+ {"chunk_id": "c2", "content_with_weight": "c", "doc_id": "d1", "docnm_kwd": "Doc One"},
+ {"chunk_id": "c3", "content_with_weight": "d", "doc_id": "d1", "docnm_kwd": "Doc One"},
+ },
+ "doc_aggs": []interface{}{
+ map[string]interface{}{"doc_id": "d1", "doc_name": "Doc One"},
+ map[string]interface{}{"doc_id": "d2", "doc_name": "Doc Two"},
+ },
+ }
+ s := &ChatPipelineService{}
+ res := s.decorateHarnessAnswer("The range claim holds [ID:1-3].", kbinfos, nil)
+
+ if strings.Contains(res.Answer, "1-3") {
+ t.Fatalf("final answer still carries the range citation: %q", res.Answer)
+ }
+ for _, want := range []string{"[ID:1]", "[ID:2]", "[ID:3]"} {
+ if !strings.Contains(res.Answer, want) {
+ t.Fatalf("final answer missing expanded citation %s: %q", want, res.Answer)
+ }
+ }
+ // Every cited chunk belongs to Doc One, so the uncited Doc Two must be
+ // filtered out of the reference (recall_docs = cited docs).
+ if aggs, _ := res.Reference["doc_aggs"].([]interface{}); len(aggs) != 1 {
+ t.Fatalf("reference doc_aggs = %#v, want the single cited doc", res.Reference["doc_aggs"])
+ }
+}
diff --git a/rag/advanced_rag/agentic_rag.py b/rag/advanced_rag/agentic_rag.py
index 79cf1082fd..53c7a6c4c9 100644
--- a/rag/advanced_rag/agentic_rag.py
+++ b/rag/advanced_rag/agentic_rag.py
@@ -217,6 +217,61 @@ def _cache_similar(
return shared / min(len(aw), len(bw)) >= _RAG_CACHE_MIN_OVERLAP
+_SLOT_CITATION_RE = re.compile(r"(?i)\[\s*ID\s*[:: ]*\s*slot\s*(\d+)\s*\]")
+_RANGE_CITATION_RE = re.compile(r"(?i)\[\s*ID\s*[:: ]*\s*([0-9]+)\s*[-–—~~]\s*([0-9]+)\s*\]")
+
+
+def _repair_slot_citation_markers(answer, slot_evidence, cite_chunk_ids):
+ """Rewrite [ID:Slot N] markers into real [ID:k] citations.
+
+ DESIGN ENHANCEMENT (Go parity: internal/service/citation.go
+ RepairSlotCitations): the model occasionally cites slot-table rows —
+ internal markers that cannot be located in the source document. Each
+ marker is rewritten to the [ID:k] of the slot's first evidence chunk that
+ is present in the citation pool (1-based kb_prompt numbering); slots
+ without pool-resident evidence have the marker removed — an internal
+ marker must not reach the answer.
+ """
+ if not answer or not slot_evidence or not cite_chunk_ids:
+ return answer
+ pos_by_id = {}
+ for i, cid in enumerate(cite_chunk_ids):
+ if cid and cid not in pos_by_id:
+ pos_by_id[cid] = i + 1
+
+ def _repl(m):
+ meta = slot_evidence.get(m.group(1)) or {}
+ for eid in meta.get("evidence_ids") or []:
+ pos = pos_by_id.get(str(eid))
+ if pos is not None:
+ return f"[ID:{pos}]"
+ return "" # unresolvable internal marker: drop
+
+ return _SLOT_CITATION_RE.sub(_repl, answer)
+
+
+def _expand_range_citation_markers(answer, pool_size):
+ """Expand range-merged citations back into individual ones.
+
+ DESIGN ENHANCEMENT (Go parity: ExpandRangeCitations): models sometimes
+ compress consecutive [ID:1][ID:2][ID:3] into [ID:1-3]; a range cannot be
+ resolved to a single source, so it is expanded when both bounds are valid
+ 1-based citation-pool positions and dropped otherwise.
+ """
+ if not answer or pool_size <= 0:
+ return answer
+
+ def _repl(m):
+ a, b = int(m.group(1)), int(m.group(2))
+ if a > b:
+ a, b = b, a
+ if a < 1 or b > pool_size:
+ return "" # out of range: unresolvable, drop
+ return "".join(f"[ID:{i}]" for i in range(a, b + 1))
+
+ return _RANGE_CITATION_RE.sub(_repl, answer)
+
+
class RAGTools:
def __init__(
self,
@@ -884,6 +939,19 @@ class RAGTools:
self.answer_sink(delta, kind == "think")
final = re.sub(r"\(\**(ID:\d+)\**\)", r"[\1]", final)
+ # DESIGN ENHANCEMENT (Go parity: internal/service/citation.go
+ # RepairSlotCitations / ExpandRangeCitations): the model sometimes
+ # cites slot-table rows ("[ID:Slot 0]") — internal markers that
+ # cannot be located in the source document — and occasionally
+ # compresses consecutive [ID:n] citations into ranges. Both forms
+ # are unresolvable and must never reach the answer.
+ final = _repair_slot_citation_markers(
+ final,
+ getattr(self, "_rag_slot_evidence", None) or {},
+ getattr(self, "_rag_cite_chunk_ids", None) or [],
+ )
+ final = _expand_range_citation_markers(final, len(getattr(self, "_rag_cite_chunk_ids", None) or []))
+
# Cache the freshly produced answer for later near-identical questions.
if question and final and not self.text_attachments_content:
self._rag_cache[question] = (final, _question_keywords(question))
diff --git a/rag/advanced_rag/agentic_rag_graph.py b/rag/advanced_rag/agentic_rag_graph.py
index 5f939e3227..d671467d84 100644
--- a/rag/advanced_rag/agentic_rag_graph.py
+++ b/rag/advanced_rag/agentic_rag_graph.py
@@ -861,6 +861,13 @@ async def _compose_answer_from_evidence(state: AgenticState, tools, token_queue:
_CITE_CHUNK_CAP = 6
cite_chunks = ranked[:_CITE_CHUNK_CAP] or all_chunks
+ # DESIGN ENHANCEMENT (Go parity: RunResponse.SlotCitations): expose the
+ # slot evidence ids and the citation-pool chunk ids so the rag tool's
+ # post-processing can rewrite unresolvable [ID:Slot N] markers into real
+ # chunk citations (see _repair_slot_citation_markers in agentic_rag.py)
+ # and expand range-merged citations (see _expand_range_citation_markers).
+ tools._rag_slot_evidence = state.get("slot_evidence") or {}
+ tools._rag_cite_chunk_ids = [str(c.get("chunk_id") or c.get("id") or "") for c in cite_chunks]
evidence_kbinfos = dict(kbinfos, chunks=cite_chunks)
evidence_blocks = kb_prompt(evidence_kbinfos, min(tools.chat_mdl.max_length, _EVIDENCE_BUDGET_TOKENS))
evidence = "\n".join(evidence_blocks) if isinstance(evidence_blocks, list) else str(evidence_blocks)
@@ -932,7 +939,11 @@ async def _compose_answer_from_evidence(state: AgenticState, tools, token_queue:
"provided evidence, those three take precedence."
)
- parts.append(f"Evidence:\n{evidence}")
+ # DESIGN NOTE: dataset names are mutable runtime data — exposed here in
+ # the untrusted evidence block (user directive), never via the system
+ # prompt's {knowledge} placeholder (trusted template content only).
+ bound = (getattr(tools, "_bound_dataset_names", "") or "").strip()
+ parts.append(f"Evidence:\n{('Bound datasets: ' + bound + '\n') if bound else ''}{evidence}")
user_content = "\n".join(parts)
_LOG.info(
diff --git a/rag/prompts/citation_prompt.md b/rag/prompts/citation_prompt.md
index 9c50c8e01f..09b14c54de 100644
--- a/rag/prompts/citation_prompt.md
+++ b/rag/prompts/citation_prompt.md
@@ -1,9 +1,9 @@
-Based on the provided document or chat history, add citations to the input text using the format specified later.
+Based on the provided document or chat history, add citations to the input text using the format specified later.
# Citation Requirements:
## Technical Rules:
-- Use format: [ID:i] or [ID:i] [ID:j] for multiple sources
+- Use format: [ID:i] or [ID:i] [ID:j] for multiple sources. NEVER merge consecutive citations into a range such as [ID:i-j] or [ID:i~j] — ranges cannot be resolved to a source.
- Place citations at the end of sentences, before punctuation
- Maximum 4 citations per sentence
- DO NOT cite content not from <context></context>
@@ -13,7 +13,7 @@ Based on the provided document or chat history, add citations to the input text
## What MUST Be Cited:
1. **Quantitative data**: Numbers, percentages, statistics, measurements
-2. **Temporal claims**: Dates, timeframes, sequences of events
+2. **Temporal claims**: Dates, timeframes, sequences of events
3. **Causal relationships**: Claims about cause and effect
4. **Comparative statements**: Rankings, comparisons, superlatives
5. **Technical definitions**: Specialized terms, concepts, methodologies
@@ -34,13 +34,13 @@ Based on the provided document or chat history, add citations to the input text
ID: 45
└── Content: The global smartphone market grew by 7.8% in Q3 2024, with Samsung holding 20.6% market share and Apple at 15.8%.
-ID: 46
+ID: 46
└── Content: 5G adoption reached 1.5 billion users worldwide by October 2024.
</context>
USER: How is the smartphone market performing?
-ASSISTANT:
+ASSISTANT:
The smartphone industry is showing strong recovery. The global smartphone market grew by 7.8% in Q3 2024 [ID:45]. Samsung leads with 20.6% market share, followed by Apple at 15.8% [ID:45]. This growth is partly driven by 5G technology, which reached 1.5 billion users worldwide by October 2024 [ID:46].
## Example 2: Causal Relationships and Predictions
@@ -114,9 +114,10 @@ ASSISTANT:
--- Examples END ---
-REMEMBER:
+REMEMBER:
- Cite FACTS, not opinions or transitions
- Each citation supports the ENTIRE sentence
- When in doubt, ask: "Would a fact-checker need to verify this?"
- Place citations at sentence end, before punctuation
- Format likes this is FORBIDDEN: [ID:0, ID:5, ID:...]. It MUST be separated like, [ID:0][ID:5]...
+- Range citations like [ID:1-3] are FORBIDDEN too. Multiple sources MUST be spelled out individually: [ID:1][ID:2][ID:3].
diff --git a/rag/prompts/generator.py b/rag/prompts/generator.py
index e88ea4edde..fa5e2a55f0 100644
--- a/rag/prompts/generator.py
+++ b/rag/prompts/generator.py
@@ -225,7 +225,12 @@ PROMPT_JINJA_ENV = SandboxedEnvironment(autoescape=False, trim_blocks=True, lstr
def citation_prompt(user_defined_prompts: dict = {}) -> str:
template = PROMPT_JINJA_ENV.from_string(user_defined_prompts.get("citation_guidelines", CITATION_PROMPT_TEMPLATE))
- return template.render() + "\n\nIMPORTANT: The example IDs above (45, 46, 78, etc.) are illustrative only. Use the actual chunk IDs from the provided knowledge blocks."
+ suffix = "\n\nIMPORTANT: The example IDs above (45, 46, 78, etc.) are illustrative only. Use the actual chunk IDs from the provided knowledge blocks."
+ # Range citations cannot be resolved to a single source; the ban is
+ # appended unconditionally so it also holds when a user overrides the
+ # template (matches Go's citationIDSuffix behavior).
+ suffix += " Cite each source INDIVIDUALLY: [ID:i][ID:j]. NEVER merge consecutive citations into a range such as [ID:i-j] or [ID:i~j]."
+ return template.render() + suffix
def citation_plus(sources: str) -> str:
diff --git a/test/unit_test/api/db/services/test_dialog_service_rag_agent_messages.py b/test/unit_test/api/db/services/test_dialog_service_rag_agent_messages.py
index 734c90e0c2..318c624fe5 100644
--- a/test/unit_test/api/db/services/test_dialog_service_rag_agent_messages.py
+++ b/test/unit_test/api/db/services/test_dialog_service_rag_agent_messages.py
@@ -222,16 +222,69 @@ def test_rag_agent_preserves_multimodal_content_parts(monkeypatch):
@pytest.mark.p2
-def test_render_reasoning_system_prompt_substitutes_date_and_knowledge():
- """The reasoning path should honor the dialog system prompt like async_chat does."""
+def test_render_reasoning_system_prompt_substitutes_date_and_knowledge(monkeypatch):
+ """The reasoning path should honor the dialog system prompt like async_chat does.
+
+ {knowledge} is trusted-template content only: mutable runtime data (bound
+ dataset names) must NOT be injected into the reasoning system prompt
+ through it. When the caller does not supply a value it renders empty —
+ the bound dataset names are exposed through the agentic graph's untrusted
+ evidence block instead (see rag/advanced_rag/agentic_rag_graph.py).
+ """
dialog = SimpleNamespace(kb_ids=["kb-1"])
prompt_config = {"system": "Role: pirate. Date: {date}. Context: '{knowledge}'."}
kwargs = {}
+ monkeypatch.setattr(
+ dialog_service.KnowledgebaseService,
+ "get_by_ids",
+ lambda ids: [SimpleNamespace(name="Pirate KB")],
+ )
rendered = dialog_service._render_reasoning_system_prompt(dialog, prompt_config, kwargs)
assert rendered.startswith("Role: pirate. Date: 2")
assert "Context: ''." in rendered
+ assert "Pirate KB" not in rendered
+
+
+def test_render_reasoning_system_prompt_caller_supplied_knowledge_wins(monkeypatch):
+ """A caller-supplied knowledge value must not be overwritten by the default."""
+ dialog = SimpleNamespace(kb_ids=["kb-1"])
+ prompt_config = {"system": "Context: '{knowledge}'."}
+ kwargs = {"knowledge": "caller evidence"}
+ monkeypatch.setattr(
+ dialog_service.KnowledgebaseService,
+ "get_by_ids",
+ lambda ids: [SimpleNamespace(name="Pirate KB")],
+ )
+
+ rendered = dialog_service._render_reasoning_system_prompt(dialog, prompt_config, kwargs)
+
+ assert rendered == "Context: 'caller evidence'."
+
+
+def test_render_reasoning_system_prompt_empty_knowledge_is_backward_compatible(monkeypatch):
+ """A dialog without datasets (or unnamed ones) still renders the template.
+
+ get_by_ids failures are swallowed: names are cosmetic and must never block
+ the prompt render.
+ """
+ dialog = SimpleNamespace(kb_ids=["kb-1"])
+ prompt_config = {"system": "Context: '{knowledge}'."}
+
+ def _boom(ids):
+ raise RuntimeError("no db in unit test")
+
+ monkeypatch.setattr(dialog_service.KnowledgebaseService, "get_by_ids", _boom)
+
+ rendered = dialog_service._render_reasoning_system_prompt(dialog, prompt_config, {})
+
+ assert rendered == "Context: ''."
+
+ no_kb = dialog_service._render_reasoning_system_prompt(SimpleNamespace(kb_ids=[]), prompt_config, {})
+ assert no_kb == "Context: ''."
@pytest.mark.p2