Multiple failures with small LLM (Ollama)

[Ehsanstp]

SelfQueryRetriever causing syntax crash and completeness_grader throwing AttributeError

• OS: Windows 11 (WSL 2, Ubuntu)
• Embedding model: sentence-transformers/all-MiniLM-L6-v2 model
• Embeddings dataset version: Reactome Release 96
• GraphDB version: reactome/graphdb:latest (Release 96)
• Generation backend: Ollama (currently: qwen2.5:0.5b)

BUGS:

A. AttributError in completeness_grader.py

This is an unhandled exception in the assess_completeness node of the external_search workflow. When the LLM output doesn't stick to the CompletenessGrade schema, the with_structured_output() returns None for the result instead of raising. The chatbot is already fully streamed to the UI before postprocess() though, regardless of the crash. What breaks is the Tavily web search fallback.

Error log:

biochat_chainlit  |   File "/app/src/external_search/completeness_grader.py", line 50, in ainvoke
biochat_chainlit  |     return {"external_search": result.binary_score}
biochat_chainlit  |                                ^^^^^^^^^^^^^^^^^^^
biochat_chainlit  | AttributeError: 'NoneType' object has no attribute 'binary_score'
biochat_chainlit  | During task with name 'assess_completeness' and id 'cfcc2097-33e8-405a-c0a7-439fda1f79d0' 

Cause:

1. Wrong format. Small LLMs return plain text or malformed JSON
2. If generation exceeds the context window, it might produce output that cannot be parsed.

B. SelfQueryRetriever Syntax Crash.

Wrong syntax generation causing error.

Error log:

Cause:

Small models generate Python-style boolean expressions ( and ) instead of the correct syntax (and_()). So the parser raises Unexpected token error.

FIX:

Both of the bugs have been investigates and fixes are ready as a PR.

BUG 1:

Adding a three-tier fallback:

• structured output
• raw text parsing and extracting the grade state
• Hard default: "No"

"No" is used as the hard default as it triggers web search rather than skipping it altogether if the grader state is unknown. Cost of false "No" is one unnecessary Tavily call.

BUG 2:

Add keyword arguments to chain_kwargs inside SelfQueryRetriever.from_llm(), this will help to reduce the generation of invalid components. But this is not guaranteed for all malformed filters. Besides prompt modification or adding solid examples in the functional format that the model can follow can help: and_(eq("gene", "TP53"), eq("synonyms_geneName", "degradation")).

Note: This won't be a problem with larger models. A similar issue was discussed in this Github issue - langchain-ai/langchain#9368

[Ehsanstp]

MODEL ISSUES:

1. The local chatbot (Ollama with qwen2.5:0.5b) produces lower quality and less consistent answers compared to live chatbot.
   This appears to stem from: Limited model capacity - 0.5B Model, Weaker instruction following, Inconsistent grounding in retrieved context.
   However, further evaluation is required to isolate whether this is due to: LLM generation quality, Retriever quality, Context or prompt formatting.

2. In some cases, retrieval results appear to be suboptimal or outright wrong, which may affect answer quality.
   Possible contributing factors: The Generation model, Embedding model (all-MiniLM-L6-v2), Reranking Failure
   Requires further controlled testing.

3. Inconsistency (Example)
   Query: “What’s the stable identifier for Lipolysis?”
   Hallucinates.

4. Query: “What’s the stable identifier for Glycolysis?”
   Expected: R-HSA-70171
   Local chatbot:

   • First run: correct
   • Second run: Incorrect (replies: R-HSA-536018)
     Suggesting inconsistencies and possible hallucination.

5. Retrieved context also appears off-target:
   Query transformers to “glycolysis pathway” with filter:
   eq(“st_id”, “R-HSA-536018”)

6. Citation Compliance issue:
   The citation requirements are often ignored and no relevant links are attached.

Positives

• Correctly identifies and describes some biological pathways, indicating that the retriever works at least partially.
• Refuses to answer questions outside its domain.
  For example “Tell me about Cricket” generates an answer that points out the question being unrelated,

Some answers generated by the local chatbot