Skip to content

Knowledgebase hybrid

DynamiqKnowledgebaseHybridSearch

Bases: Node

Composes DynamiqKnowledgebaseVectorSearch and DynamiqKnowledgebaseGraphSearch into one tool.

The two sub-searches run concurrently (via execute_async). Vector chunks and the graph's structured source documents are merged and deduplicated, then reranked against the query when a reranker is configured (otherwise all merged documents are returned). The graph's content (facts block) is always appended to the formatted output.

Source code in dynamiq/nodes/knowledgebases/knowledgebase_hybrid.py
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
class DynamiqKnowledgebaseHybridSearch(Node):
    """Composes ``DynamiqKnowledgebaseVectorSearch`` and ``DynamiqKnowledgebaseGraphSearch`` into one tool.

    The two sub-searches run concurrently (via ``execute_async``). Vector chunks and the graph's structured
    source documents are merged and deduplicated, then reranked against the query when a reranker is
    configured (otherwise all merged documents are returned). The graph's ``content`` (facts block) is
    always appended to the formatted output.
    """

    group: Literal[NodeGroup.TOOLS] = NodeGroup.TOOLS
    action_type: ActionType = ActionType.SEMANTIC_SEARCH
    name: str = "dynamiq-knowledgebase-hybrid-search"
    description: str = DESCRIPTION

    vector_search: DynamiqKnowledgebaseVectorSearch
    graph_search: DynamiqKnowledgebaseGraphSearch
    # Any RANKERS-group node with a ``{query, documents}`` -> ``{documents}`` contract (e.g. CohereReranker).
    reranker: Node | None = None
    # Per-source fetch size and the final cap on the merged result; overridable per call via input ``limit``.
    limit: int | None = None

    input_schema: ClassVar[type[DynamiqKnowledgebaseHybridSearchInputSchema]] = (
        DynamiqKnowledgebaseHybridSearchInputSchema
    )

    @property
    def to_dict_exclude_params(self):
        # Sub-nodes are serialized via their own ``to_dict`` (below) so their non-serializable runtime
        # state (e.g. HTTP ``client``) is excluded; keep them out of the parent ``model_dump``.
        return super().to_dict_exclude_params | {
            "vector_search": True,
            "graph_search": True,
            "reranker": True,
        }

    def to_dict(self, **kwargs) -> dict:
        """Serialize composed sub-nodes through their own ``to_dict`` so the node roundtrips via YAML."""
        data = super().to_dict(**kwargs)
        data["vector_search"] = self.vector_search.to_dict(**kwargs)
        data["graph_search"] = self.graph_search.to_dict(**kwargs)
        data["reranker"] = self.reranker.to_dict(**kwargs) if self.reranker else None
        return data

    def init_components(self, connection_manager: ConnectionManager | None = None):
        """Initialize the composed sub-nodes so they share the parent's connection manager."""
        connection_manager = connection_manager or ConnectionManager()
        super().init_components(connection_manager)
        for component in (self.vector_search, self.graph_search, self.reranker):
            if component is not None and component.is_postponed_component_init:
                component.init_components(connection_manager)

    def _resolve_limit(self, input_data: DynamiqKnowledgebaseHybridSearchInputSchema) -> int | None:
        """Effective limit for this run: per-call input wins over the node default."""
        return input_data.limit if input_data.limit is not None else self.limit

    def _sub_input(self, input_data: DynamiqKnowledgebaseHybridSearchInputSchema, *, graph: bool) -> dict[str, Any]:
        """Build the input dict for a sub-search, dropping keys the graph search does not accept."""
        payload: dict[str, Any] = {"query": input_data.query, "filters": input_data.filters}
        limit = self._resolve_limit(input_data)
        if limit is not None:
            payload["limit"] = limit
        if input_data.user is not None:
            payload["user"] = input_data.user
        if not graph:
            if input_data.similarity_threshold is not None:
                payload["similarity_threshold"] = input_data.similarity_threshold
            if input_data.alpha is not None:
                payload["alpha"] = input_data.alpha
        return payload

    @staticmethod
    def _sub_run_kwargs(kwargs: dict[str, Any]) -> dict[str, Any]:
        """Tracing kwargs for a sub-run: reparent it under THIS node's run.

        Sets ``parent_run_id`` to the hybrid node's own ``run_id`` so the vector/graph/reranker runs nest
        under it in the trace tree (mirrors ``Agent``). Without it the sub-runs record no parent, the tracer
        treats each as a root, and ``TracingCallbackHandler`` flushes when a sub-search finishes -- before the
        hybrid run completes. ``trace_id``/``session_id`` need no threading: they live on the shared tracer
        reached via ``config.callbacks``.
        """
        return {"parent_run_id": kwargs.get("run_id")}

    def _rerank_or_degrade(self, result: RunnableResult, documents: list[Document]) -> list[Document]:
        """Unwrap a reranker ``RunnableResult``; on failure degrade to the unranked ``documents``.

        Reranking is an optional refinement -- a reranker failure must not abort a hybrid search whose
        sub-searches already succeeded (mirrors ``KnowledgeGraphRetriever._maybe_rerank``). Cancellation
        still propagates.
        """
        if result.status == RunnableStatus.CANCELED:
            raise CanceledException()
        if result.status != RunnableStatus.SUCCESS:
            error = result.error.to_dict() if result.error else "unknown error"
            logger.warning(
                f"Tool {self.name} - {self.id}: reranker '{self.reranker.name}' failed: {error}; "
                f"returning unranked documents."
            )
            return documents
        return (result.output or {}).get("documents", documents)

    def _output_or_degrade(self, result: RunnableResult, label: str) -> dict[str, Any] | None:
        """Like ``_output_or_raise`` but degrades a failed sub-search to ``None`` (logged) instead of raising.

        Cancellation still propagates -- a canceled sub-run is never routed around.
        """
        if result.status == RunnableStatus.CANCELED:
            raise CanceledException()
        if result.status != RunnableStatus.SUCCESS:
            error = result.error.to_dict() if result.error else "unknown error"
            logger.warning(
                f"Tool {self.name} - {self.id}: {label} search failed: {error}; returning the other source."
            )
            return None
        return result.output or {}

    def _resolve_outputs(
        self, vector_result: RunnableResult, graph_result: RunnableResult
    ) -> tuple[dict[str, Any], dict[str, Any]]:
        """Resolve both sub-search outputs, tolerating a single failure. Fails only when BOTH fail.

        A failed source is logged (in ``_output_or_degrade``) and replaced by ``{}`` so the merge proceeds
        on whatever survived. Cancellation still propagates.
        """
        vector_output = self._output_or_degrade(vector_result, "vector")
        graph_output = self._output_or_degrade(graph_result, "graph")
        if vector_output is None and graph_output is None:
            raise ToolExecutionException(
                "Hybrid search failed: both the vector and graph sub-searches errored. "
                "Please analyze the error and take appropriate action.",
                recoverable=True,
            )
        return vector_output or {}, graph_output or {}

    def _dedupe_merge(self, vector_output: dict[str, Any], graph_output: dict[str, Any]) -> list[Document]:
        """Merge and deduplicate vector chunks with graph source documents (pre-rerank).

        Deduplication is by ``Document.id`` when present, else by content. Vector documents win ties so
        their scores/metadata are preserved; every document is tagged with its ``retrieval_source``.
        """
        vector_documents = [self._as_document(item, "vector") for item in (vector_output.get("documents") or [])]

        raw_graph_docs = graph_output.get("source_documents") or []
        graph_documents = [self._as_document(item, "graph") for item in raw_graph_docs]

        merged: list[Document] = []
        seen: set[str] = set()
        for document in [*vector_documents, *graph_documents]:
            key = document.id or (document.content or "")
            if key in seen:
                continue
            seen.add(key)
            merged.append(document)
        return merged

    def _finalize(
        self, documents: list[Document], graph_output: dict[str, Any], limit: int | None
    ) -> dict[str, Any]:
        """Apply the limit cap and append the graph facts block to form the final output."""
        if limit is not None:
            documents = documents[:limit]  # cap uniformly, whether or not a reranker reordered

        # Append the graph's ``facts`` (triples) -- its unique signal; source-doc text is already in the pool.
        graph_facts = (graph_output.get("facts") or "").strip()
        content = self._format_content(documents)
        if graph_facts:
            content = f"{content}\n\n--- Graph Facts ---\n{graph_facts}" if content else graph_facts

        return {"content": content, "documents": documents}

    def _merge(
        self,
        query: str,
        vector_output: dict[str, Any],
        graph_output: dict[str, Any],
        config: RunnableConfig,
        limit: int | None = None,
        **kwargs,
    ) -> dict[str, Any]:
        """Merge vector chunks with graph source documents, rerank, and append graph facts."""
        merged = self._dedupe_merge(vector_output, graph_output)
        documents = self._rerank(query, merged, config, **kwargs)
        return self._finalize(documents, graph_output, limit)

    async def _merge_async(
        self,
        query: str,
        vector_output: dict[str, Any],
        graph_output: dict[str, Any],
        config: RunnableConfig,
        limit: int | None = None,
        **kwargs,
    ) -> dict[str, Any]:
        """Async counterpart of ``_merge``: awaits the reranker so the event loop is never blocked."""
        merged = self._dedupe_merge(vector_output, graph_output)
        documents = await self._rerank_async(query, merged, config, **kwargs)
        return self._finalize(documents, graph_output, limit)

    def _rerank(
        self, query: str, documents: list[Document], config: RunnableConfig, **kwargs
    ) -> list[Document]:
        """Reorder merged documents by query relevance when a reranker is configured.

        Returns the documents unchanged (order preserved) when no reranker is set. The ``limit`` cap is
        applied by the caller so it holds for both the reranked and the unranked paths.
        """
        if not documents or self.reranker is None:
            return documents

        result = self.reranker.run(
            input_data={"query": query, "documents": documents},
            config=config,
            run_depends=kwargs.get("run_depends", []),
            **self._sub_run_kwargs(kwargs),
        )
        return self._rerank_or_degrade(result, documents)

    async def _rerank_async(
        self, query: str, documents: list[Document], config: RunnableConfig, **kwargs
    ) -> list[Document]:
        """Async counterpart of ``_rerank`` -- awaits ``run_async`` so a sync reranker is offloaded to a
        thread (via ``Node.run_async``) instead of blocking the event loop."""
        if not documents or self.reranker is None:
            return documents

        result = await self.reranker.run_async(
            input_data={"query": query, "documents": documents},
            config=config,
            run_depends=kwargs.get("run_depends", []),
            **self._sub_run_kwargs(kwargs),
        )
        return self._rerank_or_degrade(result, documents)

    @staticmethod
    def _as_document(item: Any, source: str) -> Document:
        document = DynamiqKnowledgebaseVectorSearch._to_document(item)
        metadata = dict(document.metadata or {})
        metadata.setdefault("retrieval_source", source)
        document.metadata = metadata
        return document

    def _format_content(self, documents: list[Document]) -> str:
        """Format merged documents by delegating to the vector search's numbered-source formatter."""
        return self.vector_search._format_content(documents)

    def execute(
        self, input_data: DynamiqKnowledgebaseHybridSearchInputSchema, config: RunnableConfig = None, **kwargs
    ) -> dict[str, Any]:
        config = ensure_config(config)
        check_cancellation(config)
        self.run_on_node_execute_run(config.callbacks, **kwargs)
        logger.info(f"Tool {self.name} - {self.id}: started with INPUT DATA:\n{input_data.model_dump()}")

        run_depends = kwargs.get("run_depends", [])
        sub_kwargs = self._sub_run_kwargs(kwargs)
        vector_result = self.vector_search.run(
            input_data=self._sub_input(input_data, graph=False), config=config, run_depends=run_depends, **sub_kwargs
        )
        graph_result = self.graph_search.run(
            input_data=self._sub_input(input_data, graph=True), config=config, run_depends=run_depends, **sub_kwargs
        )

        vector_output, graph_output = self._resolve_outputs(vector_result, graph_result)

        return self._merge(
            input_data.query, vector_output, graph_output, config, limit=self._resolve_limit(input_data), **kwargs
        )

    async def execute_async(
        self, input_data: DynamiqKnowledgebaseHybridSearchInputSchema, config: RunnableConfig = None, **kwargs
    ) -> dict[str, Any]:
        config = ensure_config(config)
        check_cancellation(config)
        self.run_on_node_execute_run(config.callbacks, **kwargs)
        logger.info(f"Tool {self.name} - {self.id}: started with INPUT DATA:\n{input_data.model_dump()}")

        run_depends = kwargs.get("run_depends", [])
        sub_kwargs = self._sub_run_kwargs(kwargs)
        vector_result, graph_result = await asyncio.gather(
            self.vector_search.run_async(
                input_data=self._sub_input(input_data, graph=False),
                config=config,
                run_depends=run_depends,
                **sub_kwargs,
            ),
            self.graph_search.run_async(
                input_data=self._sub_input(input_data, graph=True),
                config=config,
                run_depends=run_depends,
                **sub_kwargs,
            ),
        )

        vector_output, graph_output = self._resolve_outputs(vector_result, graph_result)

        return await self._merge_async(
            input_data.query, vector_output, graph_output, config, limit=self._resolve_limit(input_data), **kwargs
        )

init_components(connection_manager=None)

Initialize the composed sub-nodes so they share the parent's connection manager.

Source code in dynamiq/nodes/knowledgebases/knowledgebase_hybrid.py
 94
 95
 96
 97
 98
 99
100
def init_components(self, connection_manager: ConnectionManager | None = None):
    """Initialize the composed sub-nodes so they share the parent's connection manager."""
    connection_manager = connection_manager or ConnectionManager()
    super().init_components(connection_manager)
    for component in (self.vector_search, self.graph_search, self.reranker):
        if component is not None and component.is_postponed_component_init:
            component.init_components(connection_manager)

to_dict(**kwargs)

Serialize composed sub-nodes through their own to_dict so the node roundtrips via YAML.

Source code in dynamiq/nodes/knowledgebases/knowledgebase_hybrid.py
86
87
88
89
90
91
92
def to_dict(self, **kwargs) -> dict:
    """Serialize composed sub-nodes through their own ``to_dict`` so the node roundtrips via YAML."""
    data = super().to_dict(**kwargs)
    data["vector_search"] = self.vector_search.to_dict(**kwargs)
    data["graph_search"] = self.graph_search.to_dict(**kwargs)
    data["reranker"] = self.reranker.to_dict(**kwargs) if self.reranker else None
    return data