How a Prompt Is Answered — the Router and Its Search Modes
Every chat message to a GraphRAG Topic is answered in one of six modes: LOCAL, GLOBAL, HYBRID, PATH, DIRECT or ATTACHMENT. You never pick the mode yourself. A router model reads the message and decides, on every single message. This explains why two questions asked one after the other can feel very different in how they are answered.
| Mode | Used for | Searches your documents? |
|---|---|---|
LOCAL |
A question about something named: a person, a product, an error code |
Yes — entities and what is connected to them |
GLOBAL |
An overview question: "What are the main topics?" |
Yes — but only the community summaries |
HYBRID |
Questions that need details and the big picture, or when the router is unsure |
Yes — LOCAL and GLOBAL together, with follow-up questions |
PATH |
"How is A related to B?" |
Yes — follows the connections between the two |
DIRECT |
Thanks, small talk, "shorter please", "as a table" |
No — answered from the conversation |
ATTACHMENT |
A question about a file you attached to the message |
No — answered from the attached file |
The router — deciding how to answer
picks a mode"] R -->|"confidence score"| D{"Sure enough?"} D -->|"yes"| M["LOCAL, GLOBAL, PATH,
DIRECT or ATTACHMENT"] D -->|"no"| HYBRID["HYBRID
(runs both)"]
The router reports how sure it is. If it is not sure enough, the Topic uses HYBRID instead of
guessing. HYBRID is the safest choice, because it runs both LOCAL and GLOBAL. You set this limit as the
confidence threshold on the wizard’s Retrieval step (the standard is 0.6). Higher: more questions go
to HYBRID, which is slower but misses less. Lower: the router’s own choice is used more often, so
answers are faster. The fallback to HYBRID itself cannot be changed.
DIRECT has its own switch and its own threshold in the same step. It is on by default, with a threshold of 0.9. When it is on, messages like thanks or "make it shorter" are answered without a search. Lower threshold: more messages are answered directly, which is faster, but a factual question may be answered from the conversation alone. Higher: more searches, which is safer but slower. When the switch is off, every message is searched. ATTACHMENT has no switch. The router can always choose it when you attach a file.
The router never reads your documents. It only chooses the mode. A small, fast model is usually enough, because it runs on every message.
LOCAL — precise, entity-centred lookup
Use it for a question that names something specific — "tell me about X".
(entities and/or text passages)"] S1 --> S2["2 · Follow the graph
N hops from them"] S2 --> S3["3 · Gather the source
text passages"] S3 --> S4["4 · Passages + connected facts
→ answer model"] S4 --> A["Answer
+ source files"]
LOCAL is the mode closest to plain RAG. The difference is that it also pulls in what is connected to the starting points. That lets it answer "what depends on X" or "who reports to Y", where a pure similarity search only finds passages that use the same words.
You choose how LOCAL finds its starting points with the seed strategy on the Retrieval step:
| Seed strategy | Starts from | Good for |
|---|---|---|
|
Entity names that match the question |
Questions that name something |
|
Text passages that match the question |
Questions worded like your documents (steps, values) |
|
Both |
Most content. A little slower |
HYBRID’s LOCAL half uses the same seed strategy.
GLOBAL — broad, thematic synthesis
Use it for "what are the main topics here" or "summarise the risks across all of this". No single passage answers such a question.
community summaries"] S1 --> S2["2 · Map step
(take what matters from each)"] S2 --> S3["3 · Reduce step
(combine into one answer)"] S3 --> A["Answer
— no source files"]
GLOBAL answers only from community summaries. It never reads single documents when it searches. That is why a GLOBAL answer shows no source files: there is no single document to point to for an overview. It also means GLOBAL is only as good as your communities are up to date. See Configuration and functions for how to rebuild them.
HYBRID — both, plus follow-up questions
Use it when the router cannot clearly say "details" or "overview". In practice this fits many real questions, for example "what is X, and how does it fit into everything else?".
asks more if needed"] G --> F F --> DC["+ direct
passage search"] DC --> S["Final answer"] S --> A["Answer
+ source files"]
HYBRID is the most thorough mode and also the slowest. It uses LOCAL and GLOBAL as inputs, and it can ask itself follow-up questions before it writes the final answer. The number of follow-up rounds and questions is set on the Retrieval step. The most searches one answer can need is the number of rounds times the number of questions per round. Higher values give more complete answers and cost more time and model calls.
PATH — how two things are connected
Use it for "How is A related to B?". The Topic looks for chains of connections that lead from one entity to the other, and answers from them. The setting path max hops (standard: 4) limits how long a chain may be. Higher: it finds more distant links, but it is slower.
DIRECT and ATTACHMENT — no search
These modes do not search the graph.
-
DIRECT answers from the conversation so far. It fits thanks, small talk and requests to change the last answer ("shorter", "as a table"). If the router thinks a search is needed after all, the message is searched instead.
-
ATTACHMENT answers from a file you attached to the message. Attachments must be allowed in the Topic’s settings. See the GraphRAG Chat Window.
What this means in practice
-
The same question can be answered very differently, depending on how you word it. "Who is the project lead for X?" (LOCAL) and "What are the biggest risks across the project?" (GLOBAL) are answered from different data in different ways.
-
GLOBAL answers show no source files. This is expected. LOCAL, HYBRID and PATH answers cite the documents they used. Pictures can appear inside an answer when the Topic stores them.
-
If GLOBAL or HYBRID answers feel out of date, the communities have not been rebuilt since new content arrived. See Configuration and functions.
-
Every limit mentioned here (hops, entities, text passages, follow-up rounds) can be changed on the wizard’s Retrieval step. They are described field by field in Retrieval strategies.
Related
-
Models and their roles — which model routes, writes the map/reduce steps and asks the follow-up questions