Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/avnlp/agentic-med-diag/llms.txt

Use this file to discover all available pages before exploring further.

Complex clinical questions cannot be answered in a single retrieval pass. A question like “which treatment is contraindicated for a patient presenting with a given condition and a specific comorbidity?” requires chaining together at least three distinct pieces of knowledge: what the condition implies, what the comorbidity implies, and how the intersection of those two implications constrains the treatment options. Answering it correctly means issuing multiple, dependent retrieval queries — each informed by the results of the previous one — rather than a single broad lookup. This is the problem that multi-hop reasoning solves: breaking a complex question into a sequence of focused sub-queries, resolving each one in turn, and chaining the results together to arrive at a grounded final answer.

​
Sub-Query Decomposition

At the start of each reasoning cycle the system decomposes the incoming clinical question into two parallel sets of sub-queries — one for the semantic channel and one for the relational channel:
  • Semantic sub-queries — A sequence of focused natural-language questions, each targeting a specific aspect of the original query. Sub-queries may include #N back-reference markers (e.g., #1, #2) that explicitly reference the answer produced by an earlier hop. This allows later sub-queries to be conditioned on what has already been established, rather than asking each question in isolation.
  • Relational SPO triple queries — A parallel sequence of Subject–Predicate–Object queries issued to the Graph RAG backend’s knowledge graph. These queries may include Entity#N placeholders that reference specific entities resolved during earlier hops, enabling the graph traversal to follow chains of relationships across multiple steps.
Both sets of sub-queries are generated upfront from the decomposition step and then executed iteratively by their respective channel subgraphs.

​
Back-References

Back-references are the mechanism that makes multi-hop reasoning possible. Without them, each sub-query would be answered independently, with no way for a later question to build on an earlier answer.
  • #N placeholders appear in semantic sub-queries. When the pipeline resolves sub-query #2, for example, it substitutes the answer generated for sub-query #1 wherever #1 appears in the query text. This grounds the later question in the specific facts already established, rather than re-issuing a generic query.
  • Entity#N placeholders appear in relational SPO triple queries. When the pipeline processes a triple query that references Entity#1, it substitutes the entity name resolved during the first hop’s graph traversal, enabling the knowledge graph traversal to follow a connected path rather than starting from scratch.
Together, these two placeholder systems allow the pipeline to construct a chain of evidence across the knowledge graph — each hop narrowing the question based on what the previous hop found.

​
Iterative Retrieval Loop

The pipeline does not execute all sub-queries in a single pass. Instead, both the semantic channel and the relational channel run as iterative retrieval-and-reasoning loops inside their respective LangGraph subgraphs:
  1. The current sub-query (with any back-references resolved) is issued to the retrieval backend.
  2. The retrieved evidence is filtered, summarised, and used to generate a sub-answer.
  3. The sub-answer is stored and made available as a back-reference for the next sub-query.
  4. The evidence verification step evaluates whether the accumulated evidence is sufficient to move forward.
  5. If verification passes, the loop advances to the next sub-query; if it fails, the conditional expansion edge is triggered.
This loop continues until all sub-queries have been resolved or the conditional expansion step has gathered sufficient additional evidence.

​
Conditional Expansion

A conditional edge in each LangGraph channel subgraph governs whether the retrieval loop advances or expands. After the evidence verification step evaluates the current sub-answer and its supporting evidence, it emits one of two signals:
  • Sufficient — The evidence adequately supports the sub-answer. The subgraph advances to the next sub-query or, if all sub-queries are resolved, exits the loop and returns its channel output to the parent graph.
  • Insufficient — The evidence is incomplete or the sub-answer is not well-grounded. The conditional edge routes execution back to the retrieval step, triggering an additional hop with a refined or expanded query.
This mechanism ensures that the pipeline does not commit to a weakly-supported answer simply because it has exhausted its pre-planned sub-queries. It can adaptively gather more evidence when the situation requires it.

​
Example Flow

The following is an illustrative example intended to show how back-references and multi-hop retrieval work in practice. It is not a real system trace.
Consider the clinical question: “What treatment options are contraindicated for a patient with Type 2 Diabetes who also has Chronic Kidney Disease?” This question requires at least three hops to answer correctly: Hop 1 — Establish the treatment landscape for Type 2 Diabetes:
  • Semantic sub-query #1: “What are the standard pharmacological treatments for Type 2 Diabetes?”
  • Relational SPO query: (Type 2 Diabetes, treated_by, ?)
  • The semantic channel retrieves document chunks describing first-line and second-line agents (metformin, SGLT2 inhibitors, GLP-1 agonists, etc.). The relational channel returns explicit treated_by triples from the knowledge graph. Both channels generate sub-answer #1.
Hop 2 — Establish the renal implications of Chronic Kidney Disease:
  • Semantic sub-query #2: “How does Chronic Kidney Disease affect drug clearance and renal dosing? [referencing #1]”
  • Relational SPO query: (Entity#1, contraindicates, Chronic Kidney Disease) — where Entity#1 is substituted with the list of agents identified in Hop 1.
  • The semantic channel retrieves passages on renal dosing thresholds and eGFR-based contraindications. The relational channel traverses contraindicates edges for each agent from Hop 1, returning explicit contraindication triples.
Hop 3 — Synthesise contraindications for the comorbid patient:
  • Semantic sub-query #3: “Which treatments from #1 are contraindicated specifically when #2 applies?”
  • The evidence verification step checks that the accumulated evidence from Hops 1 and 2 is sufficient to support a definitive answer. If a key agent is missing from the graph traversal, the conditional expansion edge triggers an additional retrieval hop before proceeding.
Once all hops are resolved, the parent graph synthesises the semantic and relational channel outputs into a final answer that explicitly names the contraindicated agents and cites both the textual passages and the knowledge graph triples as evidence.
The fan-out/fan-in coordination between the semantic and relational channels, the conditional expansion edge, and the synthesis step are all implemented in the LangGraph state machine described in the Architecture page.

Build docs developers (and LLMs) love

Get started for freeTalk to us