docs: Documented the new no_workspace_mcp configuration property that disables workspace-local MCP configurations

config.example.yaml

@@ -134,6 +134,10 @@ enabled_mcp_servers: null        # Which MCP servers to enable by default.
                                  #     - slack
                                  # Example (comma-separated form):
                                  #   enabled_mcp_servers: github,slack,ddg-search
+no_workspace_mcp: false          # Disable loading workspace-local MCP servers from .coyote/mcp.json (default: false).
+                                 # When false (the default), Coyote merges .coyote/mcp.json from the current directory
+                                 # into the global MCP registry at startup. Workspace entries shadow global ones on
+                                 # name collision. Set to true (or pass --no-workspace-mcp) to skip this entirely.
 
 # ---- Skills ----
 # Skills are modular knowledge or capability packs the LLM can load and unload mid-conversation.
@@ -199,7 +203,7 @@ rag_chunk_size: null             # Defines the size of chunks for document proce
 rag_chunk_overlap: null          # Defines the overlap between chunks
 rag_extractor_model: null        # LLM model for graph-based entity/relationship extraction; when set, enables a graph RAG signal alongside vector and BM25
 rag_extractor_prompt: null       # Custom extraction prompt template; must contain __CHUNK__ placeholder; defaults to built-in prompt when null
-rag_graph_hops: 1                # Number of hops to expand from matched entities at query time (1 = direct neighbors; increase for denser graphs)
+rag_graph_hops: 1                # Number of hops to expand from matched entities at query time (0 = seed nodes only; 1 = direct neighbors; increase for denser graphs)
 # Defines the query structure using variables like __CONTEXT__, __SOURCES__, and __INPUT__ to tailor searches to specific needs
 rag_template: |
   Answer the query based on the context while respecting the rules. (user query, some textual context and rules, all inside xml tags)

graph.example.yaml

@@ -227,7 +227,7 @@ nodes:
     reranker_model: null       # Optional reranker for hybrid-search results
     extractor_model: null      # Optional chat model for graph-based entity/relationship extraction; enables graph RAG signal when set
     extractor_prompt: null     # Optional custom extraction prompt; must contain __CHUNK__ placeholder; uses built-in prompt when null
-    graph_hops: 1              # Graph expansion depth at query time (1 = direct neighbors; increase for denser knowledge graphs)
+    graph_hops: 1              # Graph expansion depth at query time (0 = seed nodes only; 1 = direct neighbors; increase for denser knowledge graphs)
     batch_size: 100            # Optional embedding-request batch size
     state_updates:             # {{output}} = { context: <str>, sources: [<path>, ...] }
       context: "{{output.context}}"   # writes `context` -> `reducers.context = concat`

src/cli/mod.rs

@@ -195,6 +195,9 @@ pub struct Cli {
     /// Skip discovery and application of all sbx mixins (user and built-in)
     #[arg(long, requires = "sandbox")]
     pub no_mixins: bool,
+    /// Disable loading workspace MCP servers from .coyote/mcp.json
+    #[arg(long)]
+    pub no_workspace_mcp: bool,
 }
 
 impl Cli {

src/client/azure_openai.rs

@@ -50,7 +50,7 @@ fn prepare_chat_completions(
 
     let url = format!(
         "{}/openai/deployments/{}/chat/completions?api-version=2024-12-01-preview",
-        &api_base,
+        api_base,
         self_.model.real_name()
     );
 
@@ -69,7 +69,7 @@ fn prepare_embeddings(self_: &AzureOpenAIClient, data: &EmbeddingsData) -> Resul
 
     let url = format!(
         "{}/openai/deployments/{}/embeddings?api-version=2024-10-21",
-        &api_base,
+        api_base,
         self_.model.real_name()
     );
 

src/config/app_config.rs

@@ -88,6 +88,7 @@ pub struct AppConfig {
 
     pub user_agent: Option<String>,
     pub save_shell_history: bool,
+    pub no_workspace_mcp: bool,
     pub sync_models_url: Option<String>,
 
     pub clients: Vec<ClientConfig>,
@@ -162,6 +163,7 @@ impl Default for AppConfig {
 
             user_agent: None,
             save_shell_history: true,
+            no_workspace_mcp: false,
             sync_models_url: None,
 
             clients: vec![],
@@ -238,6 +240,7 @@ impl AppConfig {
 
             user_agent: config.user_agent,
             save_shell_history: config.save_shell_history,
+            no_workspace_mcp: false,
             sync_models_url: config.sync_models_url,
 
             clients: config.clients,

src/config/memory.rs

@@ -7,7 +7,7 @@ use serde::{Deserialize, Serialize};
 
 use crate::config::{
     GIT_DIR_NAME, GITIGNORE_FILE_NAME, MEMORY_DIR_NAME, MEMORY_INDEX_FILE_NAME,
-    WORKSPACE_MEMORY_DIR_NAME, WORKSPACE_MEMORY_FILE_NAME, paths,
+    WORKSPACE_COYOTE_DIR_NAME, WORKSPACE_MEMORY_FILE_NAME, paths,
 };
 
 pub const DEFAULT_MEMORY_CAP_WITH_TOOLS: usize = 6_000;
@@ -27,7 +27,7 @@ pub enum WorkspaceMemory {
 
 pub fn discover_workspace_memory(start: &Path) -> Option<WorkspaceMemory> {
     for dir in start.ancestors() {
-        let structured = dir.join(WORKSPACE_MEMORY_DIR_NAME).join(MEMORY_DIR_NAME);
+        let structured = dir.join(WORKSPACE_COYOTE_DIR_NAME).join(MEMORY_DIR_NAME);
         if structured.join(MEMORY_INDEX_FILE_NAME).exists() {
             return Some(WorkspaceMemory::Structured {
                 workspace_root: dir.to_path_buf(),
@@ -84,8 +84,8 @@ pub fn bootstrap_workspace_memory(git_root: &Path) -> Result<PathBuf> {
 
 fn append_gitignore_entry(git_root: &Path) -> Result<bool> {
     let gitignore = git_root.join(GITIGNORE_FILE_NAME);
-    let entry = format!("{WORKSPACE_MEMORY_DIR_NAME}/{MEMORY_DIR_NAME}/");
-    let entry_no_slash = format!("{WORKSPACE_MEMORY_DIR_NAME}/{MEMORY_DIR_NAME}");
+    let entry = format!("{WORKSPACE_COYOTE_DIR_NAME}/{MEMORY_DIR_NAME}/");
+    let entry_no_slash = format!("{WORKSPACE_COYOTE_DIR_NAME}/{MEMORY_DIR_NAME}");
 
     let existing = fs::read_to_string(&gitignore).unwrap_or_default();
     let already_present = existing.lines().any(|line| {
@@ -347,7 +347,7 @@ mod tests {
         let root = temp_root("phase1");
         let workspace = root.join("workspace");
         let workspace_memory_dir = workspace
-            .join(WORKSPACE_MEMORY_DIR_NAME)
+            .join(WORKSPACE_COYOTE_DIR_NAME)
             .join(MEMORY_DIR_NAME);
         fs::create_dir_all(&workspace_memory_dir).unwrap();
         fs::write(
@@ -382,7 +382,7 @@ mod tests {
         let root = temp_root("prefer");
         let workspace = root.join("ws");
         let structured = workspace
-            .join(WORKSPACE_MEMORY_DIR_NAME)
+            .join(WORKSPACE_COYOTE_DIR_NAME)
             .join(MEMORY_DIR_NAME);
         fs::create_dir_all(&structured).unwrap();
         fs::write(structured.join(MEMORY_INDEX_FILE_NAME), "s").unwrap();
@@ -415,7 +415,7 @@ mod tests {
         let root = temp_root("indexes_only");
         let workspace = root.join("ws");
         let structured = workspace
-            .join(WORKSPACE_MEMORY_DIR_NAME)
+            .join(WORKSPACE_COYOTE_DIR_NAME)
             .join(MEMORY_DIR_NAME);
         fs::create_dir_all(&structured).unwrap();
         fs::write(
@@ -450,7 +450,7 @@ mod tests {
         let root = temp_root("drill_bodies");
         let workspace = root.join("ws");
         let structured = workspace
-            .join(WORKSPACE_MEMORY_DIR_NAME)
+            .join(WORKSPACE_COYOTE_DIR_NAME)
             .join(MEMORY_DIR_NAME);
         fs::create_dir_all(&structured).unwrap();
         fs::write(structured.join(MEMORY_INDEX_FILE_NAME), "idx").unwrap();
@@ -485,7 +485,7 @@ mod tests {
         let root = temp_root("cap");
         let workspace = root.join("ws");
         let structured = workspace
-            .join(WORKSPACE_MEMORY_DIR_NAME)
+            .join(WORKSPACE_COYOTE_DIR_NAME)
             .join(MEMORY_DIR_NAME);
         fs::create_dir_all(&structured).unwrap();
         fs::write(structured.join(MEMORY_INDEX_FILE_NAME), "idx").unwrap();
@@ -575,7 +575,7 @@ mod tests {
         let root = temp_root("walk_up");
         let workspace = root.join("ws");
         let mem_dir = workspace
-            .join(WORKSPACE_MEMORY_DIR_NAME)
+            .join(WORKSPACE_COYOTE_DIR_NAME)
             .join(MEMORY_DIR_NAME);
         fs::create_dir_all(&mem_dir).unwrap();
         fs::write(mem_dir.join(MEMORY_INDEX_FILE_NAME), "idx").unwrap();

src/config/mod.rs

@@ -143,7 +143,7 @@ const MCP_FILE_NAME: &str = "mcp.json";
 const MEMORY_DIR_NAME: &str = "memory";
 const MEMORY_INDEX_FILE_NAME: &str = "MEMORY.md";
 const WORKSPACE_MEMORY_FILE_NAME: &str = "COYOTE.md";
-const WORKSPACE_MEMORY_DIR_NAME: &str = ".coyote";
+const WORKSPACE_COYOTE_DIR_NAME: &str = ".coyote";
 const SBX_KIT_DIR_NAME: &str = "sbx-kit";
 const SBX_KIT_HASH_FILE: &str = "kit.sha256";
 const SBX_MIXIN_FILE_NAME: &str = "sbx-mixin.yaml";

src/config/paths.rs

@@ -5,7 +5,7 @@ use super::{
     GLOBAL_TOOLS_UTILS_DIR_NAME, MACROS_DIR_NAME, MCP_FILE_NAME, MEMORY_DIR_NAME,
     MEMORY_INDEX_FILE_NAME, ModelsOverride, RAGS_DIR_NAME, ROLES_DIR_NAME, SBX_KIT_DIR_NAME,
     SBX_KIT_HASH_FILE, SBX_MIXIN_FILE_NAME, SBX_MIXIN_KITS_DIR_NAME, SBX_VAULT_MIXINS_DIR_NAME,
-    SKILLS_DIR_NAME, WORKSPACE_MEMORY_DIR_NAME,
+    SKILLS_DIR_NAME, WORKSPACE_COYOTE_DIR_NAME,
 };
 use crate::client::ProviderModels;
 use crate::config::REPL_HISTORY_DIR_NAME;
@@ -118,7 +118,7 @@ pub fn global_tools_sbx_mixin_file() -> PathBuf {
 pub fn find_workspace_sbx_mixin(start: &Path) -> Option<PathBuf> {
     for dir in start.ancestors() {
         let candidate = dir
-            .join(WORKSPACE_MEMORY_DIR_NAME)
+            .join(WORKSPACE_COYOTE_DIR_NAME)
             .join(SBX_MIXIN_FILE_NAME);
         if candidate.exists() {
             return Some(candidate);
@@ -193,6 +193,24 @@ pub fn skill_file(name: &str) -> PathBuf {
     skill_dir(name).join("SKILL.md")
 }
 
+pub fn workspace_skills_dir() -> PathBuf {
+    env::current_dir()
+        .unwrap_or_default()
+        .join(WORKSPACE_COYOTE_DIR_NAME)
+        .join(SKILLS_DIR_NAME)
+}
+
+pub fn workspace_skill_file(name: &str) -> PathBuf {
+    workspace_skills_dir().join(name).join("SKILL.md")
+}
+
+pub fn workspace_mcp_config_file() -> PathBuf {
+    env::current_dir()
+        .unwrap_or_default()
+        .join(WORKSPACE_COYOTE_DIR_NAME)
+        .join(MCP_FILE_NAME)
+}
+
 pub fn validate_skill_name(name: &str) -> Result<()> {
     if name.is_empty() {
         bail!("Skill name cannot be empty");
@@ -318,7 +336,7 @@ pub fn global_memory_index_path() -> PathBuf {
 
 pub fn workspace_memory_dir_for(workspace_root: &Path) -> PathBuf {
     workspace_root
-        .join(WORKSPACE_MEMORY_DIR_NAME)
+        .join(WORKSPACE_COYOTE_DIR_NAME)
         .join(MEMORY_DIR_NAME)
 }
 
@@ -405,15 +423,21 @@ pub fn has_macro(name: &str) -> bool {
 
 pub fn list_skills() -> Vec<String> {
     let mut names = Vec::new();
-    if let Ok(rd) = read_dir(skills_dir()) {
-        for entry in rd.flatten() {
-            if let Ok(file_type) = entry.file_type()
-                && file_type.is_dir()
-                && let Some(name) = entry.file_name().to_str()
-                && entry.path().join("SKILL.md").is_file()
-                && validate_skill_name(name).is_ok()
-            {
-                names.push(name.to_string());
+    let mut seen = HashSet::new();
+
+    for dir in [workspace_skills_dir(), skills_dir()] {
+        if let Ok(rd) = read_dir(dir) {
+            for entry in rd.flatten() {
+                if let Ok(file_type) = entry.file_type()
+                    && file_type.is_dir()
+                    && let Some(name) = entry.file_name().to_str()
+                    && !seen.contains(name)
+                    && entry.path().join("SKILL.md").is_file()
+                    && validate_skill_name(name).is_ok()
+                {
+                    seen.insert(name.to_string());
+                    names.push(name.to_string());
+                }
             }
         }
     }
@@ -423,7 +447,7 @@ pub fn list_skills() -> Vec<String> {
 }
 
 pub fn has_skill(name: &str) -> bool {
-    skill_file(name).is_file()
+    workspace_skill_file(name).is_file() || skill_file(name).is_file()
 }
 
 pub fn local_models_override() -> Result<Vec<ProviderModels>> {

src/config/session.rs

@@ -261,7 +261,7 @@ impl Session {
         data["messages"] = json!(self.messages);
 
         let output = serde_yaml::to_string(&data)
-            .with_context(|| format!("Unable to show info about session '{}'", &self.name))?;
+            .with_context(|| format!("Unable to show info about session '{}'", self.name))?;
         Ok(output)
     }
 

src/config/skill.rs

@@ -117,7 +117,11 @@ impl Skill {
 
     pub fn load(name: &str) -> Result<Self> {
         paths::validate_skill_name(name)?;
-        let path = paths::skill_file(name);
+        let path = if paths::workspace_skill_file(name).is_file() {
+            paths::workspace_skill_file(name)
+        } else {
+            paths::skill_file(name)
+        };
         let content = read_to_string(&path)
             .with_context(|| format!("Failed to read skill '{name}' at {}", path.display()))?;
         Ok(Skill::new(name, &content))

src/function/mod.rs

@@ -147,10 +147,10 @@ pub async fn eval_tool_calls(
     let mut is_all_null = true;
     for call in calls {
         if let Some(msg) = ctx.tool_scope.tool_tracker.check_loop(&call.clone()) {
-            let dup_msg = format!("{{\"tool_call_loop_alert\":{}}}", &msg.trim());
+            let dup_msg = format!("{{\"tool_call_loop_alert\":{}}}", msg.trim());
             println!(
                 "{}",
-                warning_text(format!("{}: ⚠️ Tool-call loop detected! ⚠️", &call.name).as_str())
+                warning_text(format!("{}: ⚠️ Tool-call loop detected! ⚠️", call.name).as_str())
             );
             let val = json!(dup_msg);
             output.push(ToolResult::new(call, val));
@@ -870,7 +870,7 @@ impl Functions {
                 let root_dir = paths::functions_dir();
                 let tool_path = format!(
                     "{}/{binary_name}",
-                    &paths::global_tools_dir().to_string_lossy()
+                    paths::global_tools_dir().to_string_lossy()
                 );
                 content_template
                     .replace("{function_name}", binary_name)
@@ -881,7 +881,7 @@ impl Functions {
                 let root_dir = paths::agent_data_dir(agent_name);
                 let tool_path = format!(
                     "{}/{binary_name}",
-                    &paths::global_tools_dir().to_string_lossy()
+                    paths::global_tools_dir().to_string_lossy()
                 );
                 content_template
                     .replace("{function_name}", binary_name)

src/main.rs

@@ -187,7 +187,11 @@ async fn main() -> Result<()> {
     let abort_signal = create_abort_signal();
     let start_mcp_servers = cli.agent.is_none() && cli.role.is_none();
     let cfg = Config::load_with_interpolation(info_flag).await?;
-    let app_config: Arc<AppConfig> = Arc::new(AppConfig::from_config(cfg)?);
+    let mut app_config = AppConfig::from_config(cfg)?;
+    if cli.no_workspace_mcp {
+        app_config.no_workspace_mcp = true;
+    }
+    let app_config: Arc<AppConfig> = Arc::new(app_config);
     let app_state: Arc<AppState> = Arc::new(
         AppState::init(
             app_config,
@@ -559,7 +563,7 @@ async fn shell_execute(
 
             match answer_char {
                 'e' => {
-                    debug!("{} {:?}", shell.cmd, &[&shell.arg, &eval_str]);
+                    debug!("{} {:?}", shell.cmd, [&shell.arg, &eval_str]);
                     let code = run_command(&shell.cmd, &[&shell.arg, &eval_str], None)?;
                     if code == 0 && app.save_shell_history {
                         let _ = append_to_shell_history(&shell.name, &eval_str, code);

src/mcp/mod.rs

@@ -214,7 +214,53 @@ impl McpRegistry {
             spec.validate(name)?;
         }
 
-        registry.config = Some(mcp_servers_config);
+        let mut merged = mcp_servers_config;
+        if !app_config.no_workspace_mcp {
+            let ws_path = paths::workspace_mcp_config_file();
+            if ws_path.try_exists().unwrap_or(false) {
+                match tokio::fs::read_to_string(&ws_path).await {
+                    Ok(ws_content) if !ws_content.trim().is_empty() => {
+                        match interpolate_secrets(&ws_content, vault) {
+                            Ok((parsed, missing)) if missing.is_empty() => {
+                                match serde_json::from_str::<McpServersConfig>(&parsed) {
+                                    Ok(ws_config) => {
+                                        let mut loaded = Vec::new();
+                                        for (name, spec) in ws_config.mcp_servers {
+                                            match spec.validate(&name) {
+                                                Ok(_) => {
+                                                    loaded.push(name.clone());
+                                                    merged.mcp_servers.insert(name, spec);
+                                                }
+                                                Err(e) => warn!(
+                                                    "Invalid workspace MCP server '{name}': {e}. Skipping."
+                                                ),
+                                            }
+                                        }
+                                        if !loaded.is_empty() {
+                                            eprintln!(
+                                                "Loading workspace MCP servers: {}",
+                                                loaded.join(", ")
+                                            );
+                                        }
+                                    }
+                                    Err(e) => warn!(
+                                        "Failed to parse workspace MCP config: {e}. Skipping."
+                                    ),
+                                }
+                            }
+                            Ok((_, missing)) => warn!(
+                                "Workspace MCP config references missing vault secrets: {missing:?}. Skipping."
+                            ),
+                            Err(e) => {
+                                warn!("Failed to process workspace MCP config: {e}. Skipping.")
+                            }
+                        }
+                    }
+                    _ => {}
+                }
+            }
+        }
+        registry.config = Some(merged);
 
         if start_mcp_servers && app_config.mcp_server_support {
             abortable_run_with_spinner(

src/rag/graph.rs

@@ -2,12 +2,21 @@ use super::DocumentId;
 use crate::client::*;
 
 use anyhow::{Context, Result};
-use indexmap::IndexMap;
+use indexmap::{IndexMap, IndexSet};
 use petgraph::Direction;
 use petgraph::graph::NodeIndex;
 use petgraph::stable_graph::StableGraph;
+use petgraph::visit::EdgeRef;
 use serde::{Deserialize, Serialize};
-use std::collections::HashSet;
+use std::collections::{HashMap, HashSet};
+
+/// Heuristic upper bound on chunk size before warning the user that the
+/// extraction LLM call may be truncated. Not a hard limit.
+const MAX_CHUNK_CHARS: usize = 24_000;
+
+/// Maximum number of nodes the BFS may visit during a single graph_search.
+/// Keeps the synchronous traversal bounded on dense graphs.
+pub const MAX_GRAPH_NODES: usize = 500;
 
 const EXTRACTION_PROMPT: &str = r#"Extract entities and relationships from the following text chunk.
 
@@ -89,16 +98,27 @@ impl Default for KnowledgeGraph {
 
 impl KnowledgeGraph {
     pub fn merge(&mut self, doc_id: DocumentId, result: ExtractionResult) {
-        let mut chunk_nodes: Vec<u32> = vec![];
+        let mut chunk_nodes: IndexSet<u32> = IndexSet::new();
 
         for extracted in &result.entities {
             let key = extracted.name.to_lowercase();
+            let normalized_type = extracted.entity_type.to_uppercase();
             let node_raw = if let Some(&existing) = self.entity_index.get(&key) {
+                let idx = NodeIndex::new(existing as usize);
+                if self.graph.contains_node(idx) {
+                    let node = &mut self.graph[idx];
+                    if node.entity_type == "OTHER" && normalized_type != "OTHER" {
+                        node.entity_type = normalized_type;
+                    }
+                    if node.description.is_none() {
+                        node.description = extracted.description.clone();
+                    }
+                }
                 existing
             } else {
                 let entity = Entity {
                     name: extracted.name.clone(),
-                    entity_type: extracted.entity_type.clone(),
+                    entity_type: normalized_type,
                     description: extracted.description.clone(),
                 };
                 let idx = self.graph.add_node(entity);
@@ -106,7 +126,7 @@ impl KnowledgeGraph {
                 self.entity_index.insert(key, raw);
                 raw
             };
-            chunk_nodes.push(node_raw);
+            chunk_nodes.insert(node_raw);
         }
 
         for extracted in &result.relationships {
@@ -118,11 +138,14 @@ impl KnowledgeGraph {
             ) {
                 let from_idx = NodeIndex::new(from_raw as usize);
                 let to_idx = NodeIndex::new(to_raw as usize);
-                // Avoid duplicate edges
-                if !self.graph.contains_edge(from_idx, to_idx) {
+                let already_exists = self
+                    .graph
+                    .edges_connecting(from_idx, to_idx)
+                    .any(|e| e.weight().relation_type == extracted.relation_type);
+                if !already_exists {
                     let rel = Relationship {
                         relation_type: extracted.relation_type.clone(),
-                        weight: extracted.weight.unwrap_or(1.0),
+                        weight: extracted.weight.unwrap_or(1.0).clamp(0.0, 1.0),
                     };
                     self.graph.add_edge(from_idx, to_idx, rel);
                 }
@@ -158,6 +181,10 @@ impl KnowledgeGraph {
             .filter(|raw| !still_used.contains(raw))
             .collect();
 
+        if to_remove.is_empty() {
+            return;
+        }
+
         for raw in to_remove {
             let idx = NodeIndex::new(raw as usize);
             if self.graph.contains_node(idx) {
@@ -166,6 +193,57 @@ impl KnowledgeGraph {
                 self.entity_index.swap_remove(&name);
             }
         }
+
+        self.compact();
+    }
+
+    /// Rebuild the internal graph with consecutive node indices. Eliminates
+    /// the null tombstone slots that petgraph's StableGraph accumulates after
+    /// repeated `remove_node` calls, keeping serialized YAML size in check.
+    fn compact(&mut self) {
+        let mut new_graph: StableGraph<Entity, Relationship> = StableGraph::new();
+        let mut old_to_new: HashMap<u32, u32> = HashMap::new();
+
+        for &old_raw in self.entity_index.values() {
+            let old_idx = NodeIndex::new(old_raw as usize);
+            if self.graph.contains_node(old_idx) {
+                let entity = self.graph[old_idx].clone();
+                let new_idx = new_graph.add_node(entity);
+                old_to_new.insert(old_raw, new_idx.index() as u32);
+            }
+        }
+
+        for edge_idx in self.graph.edge_indices() {
+            if let Some((from, to)) = self.graph.edge_endpoints(edge_idx) {
+                let from_raw = from.index() as u32;
+                let to_raw = to.index() as u32;
+                if let (Some(&new_from), Some(&new_to)) =
+                    (old_to_new.get(&from_raw), old_to_new.get(&to_raw))
+                {
+                    let rel = self.graph[edge_idx].clone();
+                    new_graph.add_edge(
+                        NodeIndex::new(new_from as usize),
+                        NodeIndex::new(new_to as usize),
+                        rel,
+                    );
+                }
+            }
+        }
+
+        for raw in self.entity_index.values_mut() {
+            if let Some(&new_raw) = old_to_new.get(raw) {
+                *raw = new_raw;
+            }
+        }
+
+        for node_raws in self.document_entities.values_mut() {
+            *node_raws = node_raws
+                .iter()
+                .filter_map(|raw| old_to_new.get(raw).copied())
+                .collect();
+        }
+
+        self.graph = new_graph;
     }
 
     pub fn build_node_to_docs(&self) -> IndexMap<u32, Vec<DocumentId>> {
@@ -179,30 +257,79 @@ impl KnowledgeGraph {
         map
     }
 
-    pub fn expand_neighbors(&self, seed_nodes: &[u32], hops: usize) -> Vec<u32> {
-        let mut expanded: indexmap::IndexSet<u32> = seed_nodes.iter().copied().collect();
-        let mut frontier: Vec<u32> = seed_nodes.to_vec();
+    /// BFS from seed nodes with weight-decayed scoring.
+    ///
+    /// Seed node scores are provided by the caller (typically token-overlap
+    /// ratios). Each neighbor's score is `edge_weight * parent_score`, so
+    /// strongly-connected neighbors rank higher and weakly-connected ones
+    /// naturally contribute less. Traversal is capped at `MAX_GRAPH_NODES`
+    /// total nodes; the highest-scored frontier nodes are expanded first so
+    /// the budget is spent on the most relevant entities.
+    ///
+    /// Returns a map of raw node index → score (includes seed nodes).
+    pub fn expand_neighbors_scored(
+        &self,
+        seed_scores: &[(u32, f32)],
+        hops: usize,
+    ) -> IndexMap<u32, f32> {
+        let mut node_scores: IndexMap<u32, f32> = IndexMap::new();
+        for &(raw, score) in seed_scores {
+            node_scores.insert(raw, score);
+        }
+
+        let mut frontier: Vec<(u32, f32)> = seed_scores.to_vec();
+
         for _ in 0..hops {
-            let mut next_frontier: Vec<u32> = vec![];
-            for &raw in &frontier {
-                let idx = NodeIndex::new(raw as usize);
-                if self.graph.contains_node(idx) {
-                    for dir in [Direction::Outgoing, Direction::Incoming] {
-                        for neighbor in self.graph.neighbors_directed(idx, dir) {
-                            let n = neighbor.index() as u32;
-                            if expanded.insert(n) {
-                                next_frontier.push(n);
+            if node_scores.len() >= MAX_GRAPH_NODES {
+                break;
+            }
+
+            frontier.sort_unstable_by(|a, b| {
+                b.1.partial_cmp(&a.1).unwrap_or(std::cmp::Ordering::Equal)
+            });
+
+            let mut next_frontier: Vec<(u32, f32)> = vec![];
+
+            'nodes: for (raw, parent_score) in &frontier {
+                let idx = NodeIndex::new(*raw as usize);
+                if !self.graph.contains_node(idx) {
+                    continue;
+                }
+                for dir in [Direction::Outgoing, Direction::Incoming] {
+                    for edge_ref in self.graph.edges_directed(idx, dir) {
+                        let neighbor_idx = match dir {
+                            Direction::Outgoing => edge_ref.target(),
+                            Direction::Incoming => edge_ref.source(),
+                        };
+                        let neighbor_raw = neighbor_idx.index() as u32;
+                        let candidate = edge_ref.weight().weight * parent_score;
+
+                        match node_scores.entry(neighbor_raw) {
+                            indexmap::map::Entry::Vacant(e) => {
+                                e.insert(candidate);
+                                next_frontier.push((neighbor_raw, candidate));
                             }
+                            indexmap::map::Entry::Occupied(mut e) => {
+                                if candidate > *e.get() {
+                                    *e.get_mut() = candidate;
+                                }
+                            }
+                        }
+
+                        if node_scores.len() >= MAX_GRAPH_NODES {
+                            break 'nodes;
                         }
                     }
                 }
             }
+
             frontier = next_frontier;
             if frontier.is_empty() {
                 break;
             }
         }
-        expanded.into_iter().collect()
+
+        node_scores
     }
 }
 
@@ -213,6 +340,14 @@ pub async fn extract_entities(
     chunk: &str,
     prompt_template: Option<&str>,
 ) -> Result<ExtractionResult> {
+    if chunk.len() > MAX_CHUNK_CHARS {
+        warn!(
+            "Entity extraction chunk is {} chars (heuristic limit: {}); \
+             the LLM response may be truncated",
+            chunk.len(),
+            MAX_CHUNK_CHARS
+        );
+    }
     let template = prompt_template.unwrap_or(EXTRACTION_PROMPT);
     let prompt = template.replace("__CHUNK__", chunk);
     let mut messages = vec![Message::new(
@@ -250,3 +385,350 @@ pub async fn extract_entities(
     serde_json::from_str::<ExtractionResult>(&json)
         .context("Failed to parse entity extraction JSON")
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn entity(name: &str, entity_type: &str) -> ExtractedEntity {
+        ExtractedEntity {
+            name: name.to_string(),
+            entity_type: entity_type.to_string(),
+            description: None,
+        }
+    }
+
+    fn rel(from: &str, to: &str, rel_type: &str, weight: f32) -> ExtractedRelationship {
+        ExtractedRelationship {
+            from: from.to_string(),
+            to: to.to_string(),
+            relation_type: rel_type.to_string(),
+            weight: Some(weight),
+        }
+    }
+
+    fn doc(id: usize) -> DocumentId {
+        DocumentId(id)
+    }
+
+    fn extraction(
+        entities: Vec<ExtractedEntity>,
+        rels: Vec<ExtractedRelationship>,
+    ) -> ExtractionResult {
+        ExtractionResult {
+            entities,
+            relationships: rels,
+        }
+    }
+
+    #[test]
+    fn merge_deduplicates_by_lowercase_name() {
+        let mut kg = KnowledgeGraph::default();
+        kg.merge(
+            doc(0),
+            extraction(
+                vec![
+                    entity("Python", "TECHNOLOGY"),
+                    entity("python", "TECHNOLOGY"),
+                ],
+                vec![],
+            ),
+        );
+        assert_eq!(kg.entity_index.len(), 1);
+        assert_eq!(kg.graph.node_count(), 1);
+    }
+
+    #[test]
+    fn merge_chunk_nodes_no_duplicate_doc_entries() {
+        let mut kg = KnowledgeGraph::default();
+        kg.merge(
+            doc(1),
+            extraction(
+                vec![
+                    entity("Python", "TECHNOLOGY"),
+                    entity("python", "TECHNOLOGY"),
+                ],
+                vec![],
+            ),
+        );
+        let count = kg.document_entities.get(&1).map(|v| v.len()).unwrap_or(0);
+        assert_eq!(
+            count, 1,
+            "duplicate entity in one chunk should produce one doc_entity entry"
+        );
+    }
+
+    #[test]
+    fn merge_normalizes_entity_type_to_uppercase() {
+        let mut kg = KnowledgeGraph::default();
+        kg.merge(
+            doc(0),
+            extraction(vec![entity("Django", "technology")], vec![]),
+        );
+        let raw = kg.entity_index["django"];
+        assert_eq!(
+            kg.graph[NodeIndex::new(raw as usize)].entity_type,
+            "TECHNOLOGY"
+        );
+    }
+
+    #[test]
+    fn merge_promotes_type_from_other_to_specific() {
+        let mut kg = KnowledgeGraph::default();
+        kg.merge(doc(0), extraction(vec![entity("Python", "OTHER")], vec![]));
+        kg.merge(
+            doc(1),
+            extraction(vec![entity("Python", "TECHNOLOGY")], vec![]),
+        );
+        let raw = kg.entity_index["python"];
+        assert_eq!(
+            kg.graph[NodeIndex::new(raw as usize)].entity_type,
+            "TECHNOLOGY"
+        );
+    }
+
+    #[test]
+    fn merge_does_not_demote_specific_type_to_other() {
+        let mut kg = KnowledgeGraph::default();
+        kg.merge(
+            doc(0),
+            extraction(vec![entity("Python", "TECHNOLOGY")], vec![]),
+        );
+        kg.merge(doc(1), extraction(vec![entity("Python", "OTHER")], vec![]));
+        let raw = kg.entity_index["python"];
+        assert_eq!(
+            kg.graph[NodeIndex::new(raw as usize)].entity_type,
+            "TECHNOLOGY"
+        );
+    }
+
+    #[test]
+    fn merge_allows_multiple_relation_types_between_same_pair() {
+        let mut kg = KnowledgeGraph::default();
+        kg.merge(
+            doc(0),
+            extraction(
+                vec![
+                    entity("Python", "TECHNOLOGY"),
+                    entity("Django", "TECHNOLOGY"),
+                ],
+                vec![rel("Python", "Django", "implements", 0.9)],
+            ),
+        );
+        kg.merge(
+            doc(1),
+            extraction(
+                vec![
+                    entity("Python", "TECHNOLOGY"),
+                    entity("Django", "TECHNOLOGY"),
+                ],
+                vec![rel("Python", "Django", "uses", 0.8)],
+            ),
+        );
+        let from_idx = NodeIndex::new(kg.entity_index["python"] as usize);
+        let to_idx = NodeIndex::new(kg.entity_index["django"] as usize);
+        let count = kg.graph.edges_connecting(from_idx, to_idx).count();
+        assert_eq!(
+            count, 2,
+            "two different relation types should produce two edges"
+        );
+    }
+
+    #[test]
+    fn merge_deduplicates_same_relation_type() {
+        let mut kg = KnowledgeGraph::default();
+        kg.merge(
+            doc(0),
+            extraction(
+                vec![entity("A", "CONCEPT"), entity("B", "CONCEPT")],
+                vec![rel("A", "B", "uses", 1.0)],
+            ),
+        );
+        kg.merge(
+            doc(1),
+            extraction(
+                vec![entity("A", "CONCEPT"), entity("B", "CONCEPT")],
+                vec![rel("A", "B", "uses", 0.5)],
+            ),
+        );
+        let from_idx = NodeIndex::new(kg.entity_index["a"] as usize);
+        let to_idx = NodeIndex::new(kg.entity_index["b"] as usize);
+        let count = kg.graph.edges_connecting(from_idx, to_idx).count();
+        assert_eq!(
+            count, 1,
+            "same relation type should not create a duplicate edge"
+        );
+    }
+
+    #[test]
+    fn remove_documents_preserves_entity_shared_across_docs() {
+        let mut kg = KnowledgeGraph::default();
+        kg.merge(
+            doc(0),
+            extraction(
+                vec![entity("Python", "TECHNOLOGY"), entity("A", "CONCEPT")],
+                vec![],
+            ),
+        );
+        kg.merge(
+            doc(1),
+            extraction(
+                vec![entity("Python", "TECHNOLOGY"), entity("B", "CONCEPT")],
+                vec![],
+            ),
+        );
+        kg.remove_documents(&[doc(0)]);
+        assert!(
+            kg.entity_index.contains_key("python"),
+            "shared entity should survive"
+        );
+        assert!(
+            !kg.entity_index.contains_key("a"),
+            "exclusive entity should be removed"
+        );
+        assert!(
+            kg.entity_index.contains_key("b"),
+            "other doc's entity should survive"
+        );
+    }
+
+    #[test]
+    fn remove_documents_noop_on_empty_slice() {
+        let mut kg = KnowledgeGraph::default();
+        kg.merge(doc(0), extraction(vec![entity("X", "CONCEPT")], vec![]));
+        kg.remove_documents(&[]);
+        assert_eq!(kg.entity_index.len(), 1);
+    }
+
+    #[test]
+    fn remove_documents_compacts_graph() {
+        let mut kg = KnowledgeGraph::default();
+        // doc 0: A, B with an edge
+        kg.merge(
+            doc(0),
+            extraction(
+                vec![entity("A", "CONCEPT"), entity("B", "CONCEPT")],
+                vec![rel("A", "B", "uses", 1.0)],
+            ),
+        );
+        // doc 1: C only
+        kg.merge(doc(1), extraction(vec![entity("C", "CONCEPT")], vec![]));
+
+        kg.remove_documents(&[doc(0)]);
+
+        assert_eq!(kg.graph.node_count(), 1);
+        let c_raw = kg.entity_index["c"];
+        assert_eq!(
+            c_raw, 0,
+            "compacted graph should give surviving node index 0"
+        );
+        let refs = kg.document_entities.get(&1).cloned().unwrap_or_default();
+        assert_eq!(refs, vec![0u32]);
+    }
+
+    #[test]
+    fn expand_zero_hops_returns_seeds_only() {
+        let mut kg = KnowledgeGraph::default();
+        kg.merge(
+            doc(0),
+            extraction(
+                vec![entity("A", "CONCEPT"), entity("B", "CONCEPT")],
+                vec![rel("A", "B", "uses", 0.9)],
+            ),
+        );
+        let a_raw = kg.entity_index["a"];
+        let result = kg.expand_neighbors_scored(&[(a_raw, 1.0)], 0);
+        assert_eq!(result.len(), 1);
+        assert_eq!(result[&a_raw], 1.0);
+    }
+
+    #[test]
+    fn expand_one_hop_decays_score_by_edge_weight() {
+        let mut kg = KnowledgeGraph::default();
+        kg.merge(
+            doc(0),
+            extraction(
+                vec![entity("A", "CONCEPT"), entity("B", "CONCEPT")],
+                vec![rel("A", "B", "uses", 0.8)],
+            ),
+        );
+        let a_raw = kg.entity_index["a"];
+        let b_raw = kg.entity_index["b"];
+        let result = kg.expand_neighbors_scored(&[(a_raw, 1.0)], 1);
+        assert_eq!(result.len(), 2);
+        assert_eq!(result[&a_raw], 1.0);
+        let b_score = result[&b_raw];
+        assert!(
+            (b_score - 0.8).abs() < 1e-6,
+            "neighbor score should be edge_weight * parent_score = 0.8, got {b_score}"
+        );
+    }
+
+    #[test]
+    fn expand_incoming_edges_also_traversed() {
+        let mut kg = KnowledgeGraph::default();
+        // Edge goes B → A; seeding A should still discover B via incoming edge
+        kg.merge(
+            doc(0),
+            extraction(
+                vec![entity("A", "CONCEPT"), entity("B", "CONCEPT")],
+                vec![rel("B", "A", "uses", 0.7)],
+            ),
+        );
+        let a_raw = kg.entity_index["a"];
+        let b_raw = kg.entity_index["b"];
+        let result = kg.expand_neighbors_scored(&[(a_raw, 1.0)], 1);
+        assert!(
+            result.contains_key(&b_raw),
+            "B should be reachable via incoming edge from A"
+        );
+        let b_score = result[&b_raw];
+        assert!((b_score - 0.7).abs() < 1e-6);
+    }
+
+    #[test]
+    fn expand_picks_best_path_score() {
+        let mut kg = KnowledgeGraph::default();
+        // A(0.5) → C(0.9): score 0.45; B(1.0) → C(0.4): score 0.40 — A→C path wins.
+        kg.merge(
+            doc(0),
+            extraction(
+                vec![
+                    entity("A", "CONCEPT"),
+                    entity("B", "CONCEPT"),
+                    entity("C", "CONCEPT"),
+                ],
+                vec![rel("A", "C", "uses", 0.9), rel("B", "C", "uses", 0.4)],
+            ),
+        );
+        let a_raw = kg.entity_index["a"];
+        let b_raw = kg.entity_index["b"];
+        let c_raw = kg.entity_index["c"];
+        let seeds = vec![(a_raw, 0.5f32), (b_raw, 1.0f32)];
+        let result = kg.expand_neighbors_scored(&seeds, 1);
+        let c_score = result[&c_raw];
+        // Best path: B(1.0) * 0.4 = 0.4, A(0.5) * 0.9 = 0.45 → should be 0.45
+        assert!(
+            (c_score - 0.45).abs() < 1e-6,
+            "C score should reflect best path (0.45), got {c_score}"
+        );
+    }
+
+    #[test]
+    fn build_node_to_docs_maps_shared_entity_to_multiple_docs() {
+        let mut kg = KnowledgeGraph::default();
+        kg.merge(
+            doc(0),
+            extraction(vec![entity("Python", "TECHNOLOGY")], vec![]),
+        );
+        kg.merge(
+            doc(1),
+            extraction(vec![entity("Python", "TECHNOLOGY")], vec![]),
+        );
+        let n2d = kg.build_node_to_docs();
+        let raw = kg.entity_index["python"];
+        let docs = &n2d[&raw];
+        assert!(docs.contains(&DocumentId(0)));
+        assert!(docs.contains(&DocumentId(1)));
+    }
+}

src/rag/mod.rs

@@ -25,6 +25,8 @@ use std::{
 };
 use tokio::time::sleep;
 
+const BM25_SEED_SCORE: f32 = 0.5;
+
 const RAG_TEMPLATE: &str = r#"Answer the query based on the context while respecting the rules. (user query, some textual context and rules, all inside xml tags)
 
 <context>
@@ -752,14 +754,14 @@ impl Rag {
             bail!("No RAG files");
         }
 
-        if self.data.extractor_model.is_some()
-            && !new_doc_contents.is_empty()
+        if !new_doc_contents.is_empty()
             && let Some(extractor_model_id) = self.data.extractor_model.clone()
         {
             match Model::retrieve_model(&self.app_config, &extractor_model_id, ModelType::Chat) {
                 Ok(model) => match self.create_embeddings_client(model) {
                     Ok(client) => {
                         let total = new_doc_contents.len();
+                        let mut failures = 0usize;
                         for (i, (doc_id, content)) in new_doc_contents.into_iter().enumerate() {
                             progress(
                                 &spinner,
@@ -774,14 +776,21 @@ impl Rag {
                             {
                                 Ok(result) => self.data.knowledge_graph.merge(doc_id, result),
                                 Err(e) => {
-                                    debug!("Entity extraction failed for doc {doc_id:?}: {e}")
+                                    warn!("Entity extraction failed for doc {doc_id:?}: {e}");
+                                    failures += 1;
                                 }
                             }
                         }
+                        if failures > 0 {
+                            progress(
+                                &spinner,
+                                format!("Entity extraction: {failures}/{total} chunks failed"),
+                            );
+                        }
                     }
-                    Err(e) => debug!("Failed to create extractor client: {e}"),
+                    Err(e) => warn!("Failed to create extractor client: {e}"),
                 },
-                Err(e) => debug!("Extractor model not found: {e}"),
+                Err(e) => warn!("Extractor model not found: {e}"),
             }
         }
 
@@ -930,9 +939,31 @@ impl Rag {
         if kg.entity_index.is_empty() {
             return vec![];
         }
-        let query_lower = query.to_lowercase();
 
-        let mut seed_nodes: Vec<u32> = kg
+        let query_lower = query.to_lowercase();
+        let query_tokens: Vec<&str> = query_lower.split_whitespace().collect();
+        let token_count = query_tokens.len().max(1);
+
+        let score_node = |raw: u32| -> f32 {
+            let idx = NodeIndex::new(raw as usize);
+            if !kg.graph.contains_node(idx) {
+                return 0.0;
+            }
+            let entity = &kg.graph[idx];
+            let combined = format!(
+                "{} {}",
+                entity.name,
+                entity.description.as_deref().unwrap_or("")
+            )
+            .to_lowercase();
+            query_tokens
+                .iter()
+                .filter(|t| combined.contains(*t))
+                .count() as f32
+                / token_count as f32
+        };
+
+        let mut seed_scores: Vec<(u32, f32)> = kg
             .entity_index
             .iter()
             .filter(|(name, _)| {
@@ -946,52 +977,31 @@ impl Rag {
                         .any(|token| token.trim_matches(|c: char| !c.is_alphanumeric()) == name_str)
                 }
             })
-            .map(|(_, &raw)| raw)
+            .map(|(_, &raw)| (raw, score_node(raw).max(BM25_SEED_SCORE)))
             .collect();
 
-        if seed_nodes.is_empty() {
+        if seed_scores.is_empty() {
             let bm25_results = self.bm25.search(query, top_k * 2);
             'outer: for result in bm25_results {
                 if let Some(node_raws) = kg.document_entities.get(&result.document.id.0) {
-                    seed_nodes.extend(node_raws.iter().copied());
-                    if seed_nodes.len() >= top_k {
-                        break 'outer;
+                    for &raw in node_raws {
+                        seed_scores.push((raw, BM25_SEED_SCORE));
+                        if seed_scores.len() >= top_k {
+                            break 'outer;
+                        }
                     }
                 }
             }
         }
 
-        if seed_nodes.is_empty() {
+        if seed_scores.is_empty() {
             return vec![];
         }
 
         let hops = self.data.graph_hops.unwrap_or(1);
-        let expanded = kg.expand_neighbors(&seed_nodes, hops);
-
-        let query_tokens: Vec<&str> = query_lower.split_whitespace().collect();
-        let token_count = query_tokens.len().max(1);
-        let mut scored: Vec<(u32, f32)> = expanded
+        let mut scored: Vec<(u32, f32)> = kg
+            .expand_neighbors_scored(&seed_scores, hops)
             .into_iter()
-            .map(|raw| {
-                let idx = NodeIndex::new(raw as usize);
-                let score = if kg.graph.contains_node(idx) {
-                    let entity = &kg.graph[idx];
-                    let combined = format!(
-                        "{} {}",
-                        entity.name,
-                        entity.description.as_deref().unwrap_or("")
-                    )
-                    .to_lowercase();
-                    query_tokens
-                        .iter()
-                        .filter(|t| combined.contains(*t))
-                        .count() as f32
-                        / token_count as f32
-                } else {
-                    0.0
-                };
-                (raw, score)
-            })
             .collect();
         scored.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap_or(Ordering::Equal));
 
@@ -1349,11 +1359,11 @@ fn set_chunk_size(model: &Model) -> Result<usize> {
 fn set_graph_hops(default_value: usize) -> Result<usize> {
     let value = Text::new("Set graph expansion hops:")
         .with_default(&default_value.to_string())
-        .with_help_message("Number of hops to expand from matched entities (1 = direct neighbors, 2 = neighbors of neighbors)")
+        .with_help_message("Number of hops to expand from matched entities (0 = seed nodes only, 1 = direct neighbors, 2 = neighbors of neighbors)")
         .with_validator(move |text: &str| {
             let out = match text.parse::<usize>() {
-                Ok(v) if v >= 1 => Validation::Valid,
-                _ => Validation::Invalid("Must be an integer >= 1".into()),
+                Ok(_) => Validation::Valid,
+                _ => Validation::Invalid("Must be a non-negative integer".into()),
             };
             Ok(out)
         })