graphiti/TODOs.md

### 7. Suggestions for Improving Graphiti's RAG Techniques

Graphiti already employs a sophisticated RAG pipeline that surpasses standard vector search by using a dynamic knowledge graph. The following suggestions aim to build upon this strong foundation by incorporating more advanced RAG strategies to enhance retrieval accuracy, contextual understanding, and overall performance.

---

#### 1. Advanced Query Pre-processing

The current approach of combining recent messages into a single search query is effective for context. However, it can be enhanced by adding a more structured query analysis step before retrieval.

**Current State:** A recent conversation history is concatenated into a single string for searching.
*   **Source Snippet (`agent.ipynb`):**
    ```python name=examples/langgraph-agent/agent.ipynb url=https://github.com/getzep/graphiti/blob/main/examples/langgraph-agent/agent.ipynb
    graphiti_query = f'{\"SalesBot\" if isinstance(last_message, AIMessage) else state[\"user_name\"]}: {last_message.content}'
    ```

**Suggested Improvements:**

*   **Query Decomposition:** For multi-faceted questions, an LLM call could break the query down into multiple sub-queries that are executed against the graph. The results can then be synthesized.
    *   **Example:** A query like *"What non-wool shoes would John like that are available in his size (10)?"* could be decomposed into:
        1.  `search("John's preferences")` -> Retrieves facts like `John LIKES "Basin Blue"`.
        2.  `search("John's shoe size")` -> Retrieves `John HAS_SHOE_SIZE "10"`.
        3.  `search("shoes NOT made of wool")` -> Retrieves products made of cotton, tree fiber, etc.
    *   **Benefit:** This provides more targeted retrieval than a single, complex query, reducing noise and improving the relevance of retrieved facts.

*   **Hypothetical Document Embeddings (HyDE):** Before performing a semantic search, the LLM can generate a hypothetical "perfect" answer to the user's query. The embedding of this *hypothetical answer* is then used for the vector search, which often yields more relevant results than embedding the query itself.
    *   **Example:** For the query *"What makes the Men's Couriers special?"*, the LLM might generate a hypothetical fact: *"The Men's Couriers are special because they feature a retro silhouette and are made from natural materials like cotton."* The embedding of this sentence is then used to find real facts in the graph.
    *   **Benefit:** This bridges the gap between the "question space" and the "answer space" in vector embeddings, leading to better semantic matches.

---

#### 2. Enhanced Graph-Native Retrieval and Summarization

Graphiti's core strength is the graph itself. Retrieval can be made even more powerful by leveraging graph topology more deeply.

**Current State:** The system uses graph traversal for re-ranking via `center_node_uuid` and detects communities (`build_communities` in the Neptune example).

**Suggested Improvements:**

*   **Recursive Retrieval & Graph-Based Summarization:** Retrieval can be a multi-step process.
    1.  **Initial Retrieval:** Retrieve a set of initial facts/nodes as is currently done.
    2.  **Neighbor Exploration:** For the top N initial nodes, automatically retrieve their direct neighbors. This can uncover critical context that wasn't captured by the initial query.
    3.  **LLM-Powered Summarization:** Pass the retrieved sub-graph (initial nodes + neighbors) to an LLM with a prompt like: *"Summarize the key information and relationships in the following set of facts: [facts list]"*.
    *   **Benefit:** This moves beyond retrieving a simple list of facts to retrieving a *synthesized insight* from a relevant portion of the graph, which is a much more compressed and potent form of context for the final generation step.

*   **Leverage Pre-computed Communities:** The `build_communities` method is a powerful feature. These communities can be used to generate summary nodes *ahead of time*.
    *   **Implementation:** After running community detection, for each community, create a new `Summary` node. Use an LLM to generate a description for this node that summarizes the entities within that community.
    *   **Example:** A community of nodes related to "SuperLight Wool Runners" could have a summary node: *"This product family is characterized by its lightweight SuperLight Foam technology and wool-based materials."*
    *   **Benefit:** During retrieval, if the query matches this summary node, the system can retrieve a single, dense summary instead of dozens of individual product facts, leading to massive prompt compression and faster, more coherent context.

---

#### 3. Optimized Ranking and Post-processing

The final step of filtering and ranking retrieved facts is crucial for prompt quality.

**Current State:** Graphiti uses a hybrid search and likely some form of score fusion (like RRF, as hinted in `search_config_recipes`) to rank results.

**Suggested Improvements:**

*   **LLM-based Re-ranking:** After retrieving an initial set of candidate facts (e.g., the top 20-30), use a smaller, faster LLM to perform a final re-ranking pass.
    *   **Implementation:** The LLM would be prompted with the original query and each fact, and asked to output a relevance score (e.g., 1-10) or a simple "relevant/not relevant" judgment.
    *   **Benefit:** This can catch nuanced relevance that semantic or keyword scores might miss, providing a final layer of polish to the retrieved context. It is more expensive but can significantly improve the quality of the top-k results.

*   **Structured Fact Output:** Instead of returning a flat list of fact strings, the retrieval endpoint could return a structured JSON object that preserves the `Subject-Predicate-Object` nature of the facts.
    *   **Example:**
        ```json
        [
          { "subject": "John", "predicate": "IS_ALLERGIC_TO", "object": "wool" },
          { "subject": "John", "predicate": "HAS_SHOE_SIZE", "object": "10" }
        ]
        ```
    *   **Benefit:** This structured format is more easily parsed by the final generation LLM, which can be explicitly prompted to "pay attention to the relationships between subjects and objects." This is a more direct way of "prompting a graph" and can lead to more logical and accurate responses. This also helps the LLM differentiate between entities and their attributes more clearly.