@inkeep/open-knowledge 0.70.0-beta.0 → 0.70.0-beta.2

package/dist/config-schema.json

@@ -881,7 +881,10 @@
         "semantic": {
           "enabled": false,
           "baseUrl": "https://api.openai.com/v1",
-          "model": "text-embedding-3-small"
+          "model": "text-embedding-3-small",
+          "maxBatchSize": 96,
+          "maxBatchChars": 96000,
+          "docTimeoutMs": 30000
         }
       },
       "type": "object",
@@ -890,7 +893,10 @@
           "default": {
             "enabled": false,
             "baseUrl": "https://api.openai.com/v1",
-            "model": "text-embedding-3-small"
+            "model": "text-embedding-3-small",
+            "maxBatchSize": 96,
+            "maxBatchChars": 96000,
+            "docTimeoutMs": 30000
           },
           "type": "object",
           "properties": {
@@ -940,6 +946,39 @@
               "reload": "live",
               "defaultScope": "project-local",
               "description": "Optional hard cutoff: drop any \"by meaning\" match whose cosine similarity is below this value. Off by default (0) because retrieval is rank-based (the closest pages are returned regardless of absolute score) and the right cutoff is model-specific. Set it only to suppress weak matches for a specific provider/model whose cosine scale you know. Most setups should leave it unset and rely on the result-count cap."
+            },
+            "maxBatchSize": {
+              "default": 96,
+              "type": "integer",
+              "minimum": 1,
+              "maximum": 2048,
+              "scope": "project-local",
+              "agentSettable": false,
+              "reload": "live",
+              "defaultScope": "project-local",
+              "description": "Maximum number of text chunks sent in one indexing request (1–2048, default 96). The ceiling follows OpenAI's input-array limit; other providers may impose different limits. Larger requests may still be split by the character budget. Changing it preserves cached vectors and document chunking; it applies to the next indexing pass. Invalid values use the default and are reported by ok config validate and server logs."
+            },
+            "maxBatchChars": {
+              "default": 96000,
+              "type": "integer",
+              "minimum": 1,
+              "maximum": 16384000,
+              "scope": "project-local",
+              "agentSettable": false,
+              "reload": "live",
+              "defaultScope": "project-local",
+              "description": "Approximate cumulative character budget per indexing request (1–16384000, default 96000). A single chunk can exceed this budget. OpenAI separately caps total input at 300000 tokens per request; this character budget does not enforce that limit, and high values can produce rejected requests. Other providers have their own limits. Changing it preserves cached vectors and document chunking; it applies to the next indexing pass. Invalid values use the default and are reported by ok config validate and server logs."
+            },
+            "docTimeoutMs": {
+              "default": 30000,
+              "type": "integer",
+              "minimum": 1,
+              "maximum": 600000,
+              "scope": "project-local",
+              "agentSettable": false,
+              "reload": "live",
+              "defaultScope": "project-local",
+              "description": "Timeout in milliseconds for each indexing request attempt (1–600000, default 30000). A request can retry four times, and a failed grouped request can trigger individual-document requests. There is no total indexing-pass deadline: large timeouts can delay newly queued work for hours. The query timeout remains 8 seconds per attempt. Changing it preserves cached vectors and applies to the next indexing pass; an in-progress pass keeps its current timeout. Invalid values use the default and are reported by ok config validate and server logs."
             }
           },
           "additionalProperties": {}

package/dist/config.project-local.schema.json

@@ -225,7 +225,10 @@
         "semantic": {
           "enabled": false,
           "baseUrl": "https://api.openai.com/v1",
-          "model": "text-embedding-3-small"
+          "model": "text-embedding-3-small",
+          "maxBatchSize": 96,
+          "maxBatchChars": 96000,
+          "docTimeoutMs": 30000
         }
       },
       "type": "object",
@@ -234,7 +237,10 @@
           "default": {
             "enabled": false,
             "baseUrl": "https://api.openai.com/v1",
-            "model": "text-embedding-3-small"
+            "model": "text-embedding-3-small",
+            "maxBatchSize": 96,
+            "maxBatchChars": 96000,
+            "docTimeoutMs": 30000
           },
           "type": "object",
           "properties": {
@@ -284,6 +290,39 @@
               "reload": "live",
               "defaultScope": "project-local",
               "description": "Optional hard cutoff: drop any \"by meaning\" match whose cosine similarity is below this value. Off by default (0) because retrieval is rank-based (the closest pages are returned regardless of absolute score) and the right cutoff is model-specific. Set it only to suppress weak matches for a specific provider/model whose cosine scale you know. Most setups should leave it unset and rely on the result-count cap."
+            },
+            "maxBatchSize": {
+              "default": 96,
+              "type": "integer",
+              "minimum": 1,
+              "maximum": 2048,
+              "scope": "project-local",
+              "agentSettable": false,
+              "reload": "live",
+              "defaultScope": "project-local",
+              "description": "Maximum number of text chunks sent in one indexing request (1–2048, default 96). The ceiling follows OpenAI's input-array limit; other providers may impose different limits. Larger requests may still be split by the character budget. Changing it preserves cached vectors and document chunking; it applies to the next indexing pass. Invalid values use the default and are reported by ok config validate and server logs."
+            },
+            "maxBatchChars": {
+              "default": 96000,
+              "type": "integer",
+              "minimum": 1,
+              "maximum": 16384000,
+              "scope": "project-local",
+              "agentSettable": false,
+              "reload": "live",
+              "defaultScope": "project-local",
+              "description": "Approximate cumulative character budget per indexing request (1–16384000, default 96000). A single chunk can exceed this budget. OpenAI separately caps total input at 300000 tokens per request; this character budget does not enforce that limit, and high values can produce rejected requests. Other providers have their own limits. Changing it preserves cached vectors and document chunking; it applies to the next indexing pass. Invalid values use the default and are reported by ok config validate and server logs."
+            },
+            "docTimeoutMs": {
+              "default": 30000,
+              "type": "integer",
+              "minimum": 1,
+              "maximum": 600000,
+              "scope": "project-local",
+              "agentSettable": false,
+              "reload": "live",
+              "defaultScope": "project-local",
+              "description": "Timeout in milliseconds for each indexing request attempt (1–600000, default 30000). A request can retry four times, and a failed grouped request can trigger individual-document requests. There is no total indexing-pass deadline: large timeouts can delay newly queued work for hours. The query timeout remains 8 seconds per attempt. Changing it preserves cached vectors and applies to the next indexing pass; an in-progress pass keeps its current timeout. Invalid values use the default and are reported by ok config validate and server logs."
             }
           },
           "additionalProperties": {}

package/dist/constants-Bc5xIETT.mjs

@@ -0,0 +1 @@
+import"./md-manager-BGnqyqi1-BRU4tUy7.mjs";var e=`0.70.0-beta.2`;const t=`config.yml`,n=e;export{n,t};